Computers and Software

International Review on

Computers and Software (IRECOS) Contents A Modeling Tool for Dynamically Reconfigurable Systems by Sonia Dimassi, Abdessalem Ben Abdelali, Amine Mrabet, Mohamed Nidhal Krifa, Abdellatif Mtibaa

600

Efficient Hardware Implementations for Tripling Oriented Elliptic Curve Crypto-System by Mohammad Alkhatib, Adel Al Salem

609

MSG SEVIRI Image Segmentation Using a Method Based on Spectral, Temporal and Textural Features by Mounir Sehad, Soltane Ameur, Jean Michel Brucker, Mourad Lazri

618

An Effective Selection of DCT and DWT Coefficients for an Adaptive Medical Image Compression Technique Using Multiple Kernel FCM by Ellappan V., R. Samson Ravindran

628

Optimal Object Detection and Tracking Using Improved Particle Swarm Optimization (IPSO) by P. Mukilan, A. Wahi

638

Prediction Algorithms for Mining Biological Databases by Lekha A., C. V. Srikrishna, Viji Vinod

650

A Grid-Based Algorithm for Mining Spatio-Temporal Sequential Patterns by Gurram Sunitha, A. Rama Mohan Reddy

659

Online Modules Placement Algorithm on Partially Reconfigurable Device for Area Optimization by Mehdi Jemai, Bouraoui Ouni, Abdellatif Mtibaa

667

Classification of Brain Tumor Using Neural Network by Bilal M. Zahran

673

An Improved Semi Supervised Nonnegative Matrix Factorization Based Tumor Clustering with Efficient Infomax ICA Based Gene Selection by S. Praba, A. K. Santra

679

Mobility Aware Load Balanced Scheduling Algorithm for Mobile Grid Environment by S. Vimala, T. Sasikala

686

A Novel Localization Approach for QoS Guaranteed Mobility-Based Communication in Wireless Sensor Networks by P. Jesu Jayarin, J. Visumathi, S. Madhurikkha

694

Energy Efficient Data Collection Framework for WSN with Layers and Uneven Clusters by L. Malathi, R. K. Gnanamurthy

701

Trust Based Web Services with Secure Control Policies and Quality of Services Using PSO by E. S. Shamila, V. Ramachandran

710

PROCEOL: Probabilistic Relational of Concept Extraction in Ontology Learning by K. Karthikeyan, V. Karthikeyani

716

Artificial Fish Swarm Load Balancing and Job Migration Task with Overloading Detection in Cloud Computing Environments by A. Mercy Gnana Rani, A. Marimuthu, A. Kavitha

727

.

Copyright © 2014 Praise Worthy Prize S.r.l. - All rights reserved

International Review on Computers and Software (I.RE.CO.S.), Vol. 9, N. 4 ISSN 1828-6003 April 2014

A Modeling Tool for Dynamically Reconfigurable Systems Sonia Dimassi, Abdessalem Ben Abdelali, Amine Mrabet, Mohamed Nidhal Krifa, Abdellatif Mtibaa Abstract – In this paper, we suggest a tool dedicated to model systems that support the dynamic partial reconfiguration. The purpose of this tool is to provide an abstract view of the developed application. It facilitates the task of the designer by abstracting many low level details and design stages. The proposed tool enables the modeling of the application to be implemented by setting its different components and the constraints to be respected namely: reconfigurable modules, static modules, the link between the application modules, the constraints on the distribution of the area of Field-Programmable Gate Arrays (FPGA) etc. The information entered through a graphical interface of the tool will be used to generate the necessary data for Xilinx tools to pursue the various implementation stages of the dynamic partial reconfiguration technique. Copyright © 2014 Praise Worthy Prize S.r.l. - All rights reserved.

Keywords: FPGA, UML, Modeling Tool, Dynamic Partial Reconfiguration

Reconfiguration (DPR) where a part of the circuit may be adapted to changes, as needed, during the execution, while the rest of the circuit continues to operate normally [1]. In this paper, we propose a tool that supports dynamic reconfiguration for FPGAs. From a high level of abstraction and through the use of the graphical language Unified Modeling Language (UML), we will be able to deal with a dynamic reconfiguration without any awareness of the underlying hardware. Through this tool, we offer a model that allows the modeling of the application to be implemented by setting up its different components, and the constraints to be respected. This paper is structured following eight parts. After the introduction, we give an overview of the different modes of the FPGA reconfiguration; in the third section, we describe the different methods of the dynamic partial reconfiguration. In the fourth one, we present the basic ideas and modeling steps of our tool. The fifth section shows how concrete designs can be realized on the basis of the use of the case diagram and the class diagram. In the sixth section, we present the proceeding steps of our modeling tool. A sample application is presented in the seventh section. Finally, we end up with a conclusion.

Nomenclature CLB DPR DR EAPR FPGA GMF HDL I/O PAR SR TDR UCF UML VGA VHDL

Configurable Logic Block Dynamic Partial Reconfiguration. Dynamic Reconfiguration Early Access Partial Reconfiguration Field Programmable Gate Arrays Graphical Modelling Framework Hardware Description Language Input / Output Place and Route Static Reconfiguration Total Dynamic Reconfiguration User Constraint File Unified Modeling Language Video Graphics Array VHSIC Hardware Description Language

I.

Introduction

Embedded systems are growing rapidly. They play an important role in many applications' fields such as mobile phones, automobiles, and digital assistances, etc. Modern embedded systems and particularly real-time embedded systems need to carry out very tough tasks and require a big computing capacity. The FieldProgrammable Gate Arrays represent one of the most popular implementation options. The reconfiguration of FPGAs can be done statically or dynamically to allow the exchange of the design parts during the given run time. The dynamic reconfiguration allows the exchange of application tasks when the system is running. Thus, the processing will take place in space and time. The current FPGA technology offers the possibility of a total Dynamic Reconfiguration and a Dynamic Partial

II.

FPGA Reconfiguration Modes

In recent years, technological advances have allowed the emergence of a new type of architecture called the reconfigurable architecture. The basic idea of this architecture is to provide the designers with the flexibility of a programmable architecture and a temporal performance of a dedicated device.

Manuscript received and revised March 2014, accepted April 2014


600

Sonia Dimassi et al.

A circuit is said to be configurable when its functionality is not predefined during the manufacturing process but may be specified in a later configuration. The FPGA is an example of a fine grain of reconfigurable architectures. An FPGA is an integrated circuit designed to be configured by a customer or a designer after being manufactured. It is composed of a Configurable Logic Block (CLB) regularly distributed and interconnected through connections that form a routing matrix [2]. An application for a static reconfiguration includes logical elements in a static manner throughout its duration, while an application for a dynamic reconfiguration uses a dynamic allocation reorganizing the material during the execution [3]. For the Dynamic Reconfiguration (DR), each application consists of several configurations for each FPGA, unlike the applications using the Static Reconfiguration (SR) that configure the FPGA permanently as shown in Fig. 1. In the DR case, FPGA can be reconfigured several times during the execution of an application.

These partitions are not exclusive since many of them can be active at the same time, whereas with a total dynamic reconfiguration just one partition is active at a given time. All these partitions or operations may be considered as independent modules that are loaded into the FPGA on demand. It is important to note that many of these partitions can be charged simultaneously and each partition can hold an arbitrary number of FPGA resources [2], [5], [6], [7].

III. Dynamic Partial Reconfiguration Design Methods For the implementation of the DPR, the manufacturer Xilinx has proposed various design methods, depending on the FPGA version and tools. Xilinx suggested the difference-based partial reconfiguration in 2002 and the module-based partial reconfiguration in 2004, which are two basic styles of the dynamic reconfiguration on a same FPGA [8]. In March 2006, Xilinx introduced the Early Access Partial Reconfiguration (EAPR) design flow along with the introduction of slice based bus macros. III.1. Difference-Based Partial Reconfiguration

Fig. 1. Static Reconfiguration

Difference-based partial reconfiguration is used for introducing small changes to design parameters such as logic equations, filter parameters, and Input /Output (I/O) standards. If there are large changes in the functionality or structure of a design, this design flow is not recommended, for example, changing an entire algorithm. When there are sizable changes or the routing has to be modified, the recommended flow is to start from the Hardware Description Language (HDL). The difference-based partial reconfiguration design flow allows a designer to introduce small logic changes using FPGA_Editor and generates a bitstream that programs only the difference between the two versions of the design. Switching the configuration of a module from one implementation to another is very quick, because the bitstream differences can be much smaller than the entire bitstream device [9].

Two basic approaches exist for implementing the applications using the dynamic reconfiguration: a total or partial dynamic reconfiguration. Both approaches use various configurations to achieve a single application, and both reconfigure the FPGA during the execution of the application. The difference between these two approaches lies in the hardware allocation. The Dynamic Partial Reconfiguration (DPR) of the FPGA, which is illustrated in Figure 2, is an approach that is more flexible than the Total Dynamic Reconfiguration (TDR). DPR allows reconfiguring logical subsets of a local or selective manner during the execution. This flexibility enables a finer allocation of hardware resources than that allowed by the TDR. Indeed, only logical elements needed are loaded into the material resources. The loading time of the configurations will be reduced and there will be a more efficient use of the hardware. The partial dynamic reconfiguration also allows the possibility of not reconfiguring the dynamic parts of the system and the static functions remain invariant throughout the FPGA implementation [4]. The design of applications with a dynamic partial reconfiguration is based on a functional partitioning.

III.2. Module-Based Partial Reconfiguration The module-based partial reconfiguration is a constructive approach in which each component has been implemented separately. This design flow is used to modify a well-defined part of the FPGA. For each module, the designer generates the bitstream configuration from a behavioural HDL description and through the phases of synthesis that are: mapping, placement and routing, independently of the other modules of the same application. Some of these modules may be reconfigured and others are static [10]. The module-based partial reconfiguration approach has a serious limitation: with Virtex II architecture, the

Fig. 2. Dynamic Partial Reconfiguration


International Review on Computers and Software, Vol. 9, N. 4

601


modules occupy the full height of the device, so the connectivity and topology have been limited to 1D structure [11]. The resources located inside the modules could not be shared with other modules and no routing is permitted by the module. Another major drawback is about the access to the I/O resources.

module must be connected through the bus macro [15], and defined by the top-level module. In this design, the control manager is implemented by the based design and the static modules cannot contain any clock or reset primitives. Similarly to the static modules, the partial reconfiguration modules cannot contain global clock signals, but may use those from the Top-level module. The reconfigurable module instantiation located in the Top-level module, represents the component name and port configuration of each module, when multiple reconfigurable modules use the same reconfigurable area.

III.3. Early Access Partial Reconfiguration This approach is an improvement and updating of the existing modular design flow. Partially reconfigurable areas could be defined as a black block that does not extend over the full height of chips [12]. This provided, to the family of Virtex 4 and others, the possibility of reconfiguring rectangular areas instead of whole columns, unlike previous technologies using the modular design flow, which physically lacked this ability. Another improvement was that the signals passing through partially reconfigurable regions without interacting do not need to be penetrated by the bus macros [13]. This significantly reduces the time constraints, improves performance and simplifies the design process. The design flow of the early access partial reconfiguration is illustrated in Figure 3 and described in details below [4], [7], [14].

III.3.2. Define Design Constraints The next step is to define the design constraints, which include the area group, reconfiguration mode, timing constraint and location constraints. In this step, the reconfigurable and static modules are specified by the area group constraint in the top-level module. So, the reconfiguration mode constraint is only applied to the reconfigurable group. For every pin, clocking primitive, and bus macro in the Top-level design, location constraints must be provided. Bus macros are located between the partial reconfiguration region and the base design.

HDL Design description

III.3.3. Implement Base Design The constraints file must be created before the implementation of the static modules. To implement the partial reconfiguration modules, we need the information generated by implementing the base design. Translate; map and Place and Route (PAR) are performed before the base design implementation.

Define Design constraints

Implement Base Design

III.3.4. Implement Partial Reconfigurable Modules

Implement Partial Reconfigurable modules

Once the implementation of modules is carried, each partial reconfigurable module must be implemented alone and follows the three following steps: translate, map and Place and Route.

Merge

III.3.5. Merge

Download

The last step is the merge phase in which the design is accomplished from each partial reconfigurable module and the base design. In this phase, many partial bitstreams, for each partial reconfigurable module and initial full bitstreams, are produced in order to configure the FPGA [14].

Fig. 3. EAPR design flow

III.3.1. Hardware Description Language (HDL) Design The initial steps in the EAPR design flow are analogous to the initial steps in the standard modular design flow. Indeed, the Top-level module contains only I/O instantiations, clock primitives, static module instantiations, partial reconfiguration module instantiations, and signal declarations and does not contain any logic. The based design and each partial reconfiguration

IV.

Basic Ideas and Modeling Steps

In this section, we focus on the definition of a modeling diagram of applications for a dynamic partial reconfiguration. This based on the concepts and characteristics of



602


reconfigurable systems that we presented previously. Our model is dedicated to model applications for a total or partial dynamic reconfiguration; it takes into account the Xilinx FPGA, and the principle of design flow EAPR of systems with a dynamic partial reconfiguration. However, the designer should not have a deep knowledge about the details of development stages of the low-level of the dynamic partial reconfiguration. Our contribution is that, referring to the design flow EAPR, our tool is at the start of the design process. The flow of our tool is illustrated in Figure 4. Different inputs data for the design flow EAPR are generated through the model introduced in our tool.

about the location of resources required for each task, while taking into account the forbidden zones. The definition of zones also allows differentiating between a reconfigurable zone and a static zone [5], [7]. Fig. 5 shows an example of partitions containing zones.

Fig. 5. Example of partitions

These zones will be placed along the constraints imposed by the manufacturer's surface. The EAPR method provides the possibility for Virtex4 and others to reconfigure rectangular zones instead of entire columns, unlike previous technologies. As EAPR design method is based on modules, the node module presents the main element of our model as shown in Figure 6. Some of these modules may be reconfigurable or static. Each module must be referenced to a HDL description; reconfigurable modules must be assigned to the reconfigurable zones and static modules must be assigned to zones defined as static zones.

FPGA

Fig. 4. EAPR design flow

First, we need to define the concepts of the system to model and specify the different needs and constraints necessary to abstract the characteristics of the system. In our case, the concepts and system characteristics are defined based on the principle of the dynamic partial reconfiguration and the design flow of systems with a dynamic partial reconfiguration on FPGA. In order to model an application, we need to define a model with the necessary elements to provide an abstract view of the application. In this model, the FPGA will be presented using its physical constraints and resources (processor). A virtual image of the FPGA will be presented through a node of our model called partition. It gives different possible configurations of the FPGA and placement constraints [6], [7]. In our model, we also define the so-called ''zone''. The definition of zones is necessary to have an idea

Bus Macro

Module statique

PR Module A PR Module A1 PR Module A2

Fig. 6. The architecture design based on EAPR

To maintain a proper communication between different application modules, a special macro bus must be used as illustrated in Fig. 6 [16]. In fact, the dynamic partial reconfiguration requires that the signals used by the reconfigurable modules to communicate use resources of fixed interconnections. These interconnected resources should not be modified when a module is changed. The signals are shown in our model by interconnections, and they are introduced by the designer graphically. So a node called bus-macro is needed in our model through a VHSIC Hardware Description Language (VHDL) code to ensure communication between two modules (static module & reconfigurable module). To perform the tasks of an application for a dynamic partial reconfiguration, the designer must take into



603


account the order of tasks to be performed according to the scenario of the application to implement. In this context, we introduced a node called sequence to handle different scenarios. This node allows the designer to choose the order of execution of the modules.

V.

UML Design of the Graphical Editor

The proposed tool should allow the modeling of the application to be implemented by setting its different elements, as well as constraints to be respected and preferences to be ensured namely: reconfigurable modules, static modules, the link between application modules, the constraints on the distribution of the surface of the FPGA, etc. The Unified Modeling Language is attracting more and more attention of Embedded Systems designers [17]. To proceed with the creation of our tool, an UML Meta model must be designed to present a simplification of reality to better understand the different needs of the graphical editor to develop [18]. This editor should allow the designer or any other user of our tool to identify one or more FPGA chips, available to implement its modeled application. V.1.

Use Case Diagrams

The use case diagrams are UML diagrams used to give an overview of the functional behaviour of the software system [19]. A use case is a discrete unit of interaction between a user and a system. In what follows, we present two use case diagrams, the first allows the designer to define devices and processors and the second use case allows the designer (actor in the diagram) to model its application. In Fig. 7, we present the use case diagram of the first graphical editor. This diagram represents the functionality that the system must take into account. The designer can define devices and processors and their characteristics. In this graphical editor, a processor is presented by a forbidden zone.

Fig. 8. Use case diagram 2

Fig. 8 shows the second use case diagram for the designing of our tool for systems with a partial dynamic reconfiguration. It is dedicated to a second graphical editor allowing the designer to create the modules (static or reconfigurable modules), the Top module of the application, sequences, the bus-macro and the different necessary zones for the implementation. Also in this use case, the designer must rely on HDL descriptions for the different modules of the application, the User Constraint File (UCF) of the Top module and ports that maintain communication between modules through bus-macro connections. V.2.

Class Diagram

UML class diagram is a type of static structure that shows the classes of the system, their attributes, operations and the interrelationships among objects. The class diagram is used for a variety of purposes, including conceptual modeling / domain modeling and detailed design. Fig. 9 shows the class diagram of the first editor that allows the creation of a library of different FPGAs with their details. Each processor must be presented in the interior of the device. The number of processors presented in a FPGA depends on the physical characteristic of the device. Figure 10 below shows the second class diagram. This diagram contains the different classes that are needed: the module class, which contains the behavioural

Fig. 7. Use case diagram 1



604


description in VHDL, the Port class, which defines the various ports of the application to model, these ports can be of input or output type and they must be presented within the modules, the Top class, which contains the behavioural description and the UCF that must be predefined according to the input/output pins of the FPGA. Also, the zones are presented by the class Area whose attributes are the dimensions and coordinates of the area, etc.

very appropriate for complex designs. With instances, we have now the possibility to cover a wider area of reconfigurable system designs.

Fig. 10. Detailed Class diagram

VI.

Realization of the Tool

In this section, we present the steps of our tool's realization on the Eclipse environment using the plug-in Graphical Modelling Framework (GMF). GMF is an important new framework for standard graphical modeling editors. Indeed, it allows quick and intuitive development of the graphical interface model. GMF generates a graphical editor from an Ecore model (Fig. 10), so we must first have a model.

Fig. 9. Class diagram 1

A "Top" is a class containing the maximum number of modules that can be present on a FPGA simultaneously. In addition to including the modules, Tops also contain overall pieces of information like the clock which is necessary for the execution. We can set the modules of each top either manually or automatically as generated by the connection of the modules. A graph can be used to view the modules and their interdependencies, so that the device can be able to authenticate the Tops. Moreover, the physical restrictions of modules present on the FPGA at the same time. Class instance, by definition, shows a snapshot of the detailed state of a system at a point in time. The module instance is an instantiation of the class “Module”. Module instances have become concrete after being added to the class "Top", or in other words they acquired a physical location. There can be many instances for single module systems if used in the same context and if they do not change their geometrical location. Instances of different Tops are identical. It is the case for all static modules, which only have matching instances. Also, instances of modules that are located in different positions at run time should stay alienated. Even if the insertion of module instances can complicate simple designs, like the designs that target the interchange of small parts of algorithms only, they are

VI.1. GMF The subproject GMF supports the creation of graphical editors to manipulate such a system in the form of diagrams. Defining a graphical editor it returns to develop:  A Domain Model that contains all the characteristics of the field;  A Graphical Definition Model;  A definition of the tools that will be present in the palette of the editor;  A Model linking these three models and the last model that allows the generation of the editor. From the model in Figs. 9 and 10, the GMF can generate the necessary elements of a graphical editor for these models. The different steps of the process for generating the graphical editor are as follows: The first step is to generate the Graphical Definition Model; the second one generates the Tooling Definition Model; the third allows generating the Mapping Model and the last one engenders the GMFGen model. The general structure of the project is shown in Fig. 11.



605


reconfigurable region that can contain two reconfigurable modules and a dedicated area for static modules. The palette on the right has enabled us to achieve this design using nodes and connections. Modules and HDL files can be inserted in the system with the assistance of some basic views. Even the Tops which have been designed in the exterior can be included.

Develop a Domain

Model * .ecore Create Project GMF

Develop Mapping model

Develop a Graphical Definition Model

* .gmfmap

* .gmfgraph Develop a Tooling Definition Model

Create Generator Model * .gmftool * .gmfgen Generate the

diagram

Fig. 11. Flow chart of GMF in Eclipse

VII.

A Modeling Example with the Tool

To illustrate the use of our tool, we perform a case study of an application of a dynamic partial reconfiguration. This application allows generating a Video Graphics Array (VGA) signal and displaying different forms of bands on a screen connected to the chip: horizontal bands or vertical bands. Each of the two display modes corresponds to a partial configuration (Fig. 12). For this, we need to create a project with our tool to illustrate the importance of it in the description of each node in the model, and the various relationships between diagram elements, which allow to provide an abstract view of the application and to facilitate the task of the designer. VGA output

Fig. 13. Tool_Device

Static Module

Virtex-4

Fig. 14. Tool_Model

Horizontal display

There exist views, suggested to the users to construct the Tops. T he Tops allow users to define many modules that a Top should contain. Moreover, the environment can include modules and allows the establishment of the structure of static communication, in order to accurately incorporate the dynamically reconfigurable parts throughout the system. Lastly, the tool allows the association of modules and Tops with their interdependencies. Hence, we offer multiple views to express priority restrictions, etc. Fig. 15 shows the information introduced by our tool. This information will be used to generate the necessary data for Xilinx tools, to pursue the various stages of implementation of the dynamic partial reconfiguration technique. In the context of the tool, we have the tree structure of files [16] to save the HDL and UCF files. Object diagrams of the file structure of the underlying operating system verify the theory of storage, so that the Xilinx data can be easily found by users.

Vertical display

Fig. 12. Example of application

As there are two Meta models, our project consists of setting up two dependent subprojects, which are the Tool_Device and the Tool_Model.  Tool_Device: sub-project that allows generating the graphical editor for devices (Fig. 13).  Tool_Model: sub-project that generates the graphical editor for modeling applications with dynamic partial reconfiguration (Fig. 14). As shown in Fig. 14, our FPGA is Virtex-II Pro, the partition is represented by a sequence consisting of a Copyright © 2014 Praise Worthy Prize S.r.l. - All rights reserved


606


[6]

[7]

[8] [9]

[10]

[11]

[12] [13]

[14]

[15]

Fig. 15. The information entered by the tool

VIII. Conclusion

[16]

We presented in this article an approach to model dynamic reconfigurable systems. This methodology is used through a tool based UML. Our tool has been carried out under the Eclipse plug-in GMF that helps create graphical editors. We also illustrated the use of our tool through a case study of an application generating a VGA signal. This illustration is presented by the description of each node in the model diagram of the application (FPGA static modules, reconfigurable modules, sequence, partition, etc.) and the different relationships between them. The information entered through the graphical interface of the tool for modeling the application, will be used to generate the data necessary for the Xilinx tools to pursue the various stages of the implementation of the dynamic partial reconfiguration technique.

[17]

[18] [19]

Authors’ information Sonia Dimassi has received her degree in Microelectronic and her master in Computer science, option: software engineering from the “University of Quebec at Montreal (UQAM)”, Canada, respectively in 1998 and 2006. She is currently working as an assistant professor in the department of computer science at the High Institute of Technological Studies of Sousse, Tunisia. She is at present preparing her Ph. D. in the “National School of Engineering of Monastir ENIM)”, Tunisia.Her research interests include methods and development tools for reconfigurable architectures and partitioning algorithms HW/SW to optimize logic area for System on Programmable Chip (SOPC).

References [1]

[2]

[3]

[4] [5]

Bouraoui Ouni, Ramzi Ayadi, Abdellatif Mtibaa, Combining temporal partitioning and temporal placement techniques for communication cost improvement. Advances in Engineering Software 42(7): 2011, pp. 444-451. Bouraoui Ouni, Ramzi Ayadi, Abdellatif Mtibaa, Partitioning and scheduling technique for run time reconfigured systems. IJCAET 3(1): 2011, pp. 77-91. Xilinx XAPP290 (v2.0) December 3, 2007 Grégory Mermoud, A Module-Based Dynamic Partial Reconfiguration tutorial. Logic Systems Laboratory École Polytechnique Fédérale de Lausanne. 2004. S. Meftali, I.R. Quadri et J.L. Dekeyser, A model based design flow for dynamic reconfigurable fpgas. International Journal of Reconfigurable Computing, INRIA LILLE NORD EUROPE LIFL - University of Lille - CNRS, Lille, France. 2009. A. Donlin, Designing Embedded Processors - A Low Power Perspective, Chapter 22: Applications, design tools and low power issues in fpga reconfiguration, book, 2007. Xilinx Corporation, Early access partial reconfiguration user guide, ug208 v1.1, www.xilinx.com, 2006. Robert Ogor, Cours Modélisation avec UML, Mai 2003 Richard C. Gronback ECLIPSE MODELING PROJECT A DomainSpecific LanguageToolkit, 2003. Wang Lie, Wu Feng-yan, Dynamic partial reconfiguration in FPGAs. Third International Symposium on Intelligent Information Technology Application Department: Computer Science and Electronic Information, Guangxi University LYan09Nanning, China, 2009. Mohamed Nidhal KRIFA, Abdessalem BEN ABDELALI et Abdellatif MTIBAA, Conception D’un Système A Base De FIR Filtre Adaptatif En Utilisant La Reconfiguration Dynamique Partielle, SETIT 2012. Amine MRABET, Mohamed Nidhal KRIFA, Sonia DIMASSI, Abdessalem Ben Abdelali, Abdellatif Mtibaa, Modélisation des SOPC à Reconfiguration Dynamique Partielle. International Conference on Embedded Systems in Telecommunications and Instrumentation (ICESTI'12): NOVEMBER 5, 6 and 7, 2012 Annaba Algeria. Boutekkouk, F., Benmohammed, M., Control/data driven embedded systems high level modeling, formal verification and simulation, (2009) International Review on Computers and Software (IRECOS), 4 (1), pp. 73-87. Xilinx, Two Flows for Partial Reconfiguration: Module Based or Difference Based, Application Note 290, Xilinx, 2004. Jakimi, A., Elkoutbi, M., A new approach for UML scenario engineering, (2009) International Review on Computers and Software (IRECOS), 4 (1), pp. 88-95.

Legat, U., Biasizzo, A., Novak, F., Partial runtime reconfiguration of FPGA, applications and a fault emulation case study, (2009) International Review on Computers and Software (IRECOS), 4 (5), pp. 606-611. Bouraoui Ouni, Optimization Algorithms for Reconfigurable FPGA Based Architectures: FPGA, design flow, reconfigurable architectures, System on Programmable chip LAP LAMBERT Academic Publishing, 200 pages, 31 May 2012. P. Lysaght et J. Dunlop. Dynamic reconfiguration of field programmable gate Array, International workshop on Field Programmable Logic and Application. Oxford Engleterre, Septembre, 2003. Xilinx Corporation, Partial Reconfiguration User Guide, www.xilinx.com, 2010. Abdellatif Mtibaa, Bouraoui Ouni, Mohamed Abid , An efficient list scheduling algorithm for time placement problem, Computers & Electrical Engineering (CEE) 33(4): 2007, pp. 285-298.

Abdessalem Ben Abdelali has received his degree in Electrical Engineering and his DEA in industrial informatics from the National School of Engineering of Sfax (ENIS), Tunisia, respectively in 2001 and 2002. He received his Ph.D from ENIS and Burgundy University (BU), France, in 2007. Since 2008 he has been working as an Assistant Professor in digital embedded electronic at the High Institute of Informatics and



607


Mathematics of Monastir (ISIMM). His current research interests include reconfigurable architectures, dyn and hardware implementation of image and video processing applications. Amine Mrabet received his M. Sc. Degree in Computer Science “Intelligent and Communication Systems, Option: Embedded Micro-electronics Systems” from the Engineering School of Sousse, Tunisia in 2012. He is currently preparing his Ph. D. degree at the Electronic and Micro-Electronic Laboratory, Faculty of Sciences of Monastir and the Engineering School of Tunis. His research interests include efficient implementation of cryptographic applications on FPGA. Mohamed Nidhal Krifa Received his degree in Electrical Engineering from the National School of Engineering of Gabes (ENIG), Tunisia, in 2004, and his master degree in microelectronics from the faculty of sciences of Monastir, Tunisia, in 2007. He is currently a PhD student at the National School of Engineering of Monastir (ENIM), Tunisia. Since 2011 he has been working as an Assistant Professor in electronics at the high Institute of Applied Science and Technology of Gafsa (ISSATG). His current research interests include reconfigurable architectures and dynamic partial reconfiguration. Abdellatif Mtibaa is currently a Professor in Micro-Electronics and Hardware Design with the Electrical Department at the National School of Engineering of Monastir and the Head of Circuits Systems Reconfigurable-ENIMGroup at Electronic and Microelectronic Laboratory. He received his Diploma in Electrical Engineering in 1985 and his PhD in Electrical Engineering in 2000. His current research interests include system on programmable chip, high level synthesis, rapid prototyping and reconfigurable architecture for real-time multimedia applications. He has authored/co-authored over 100 papers in international journals and conferences. He served on the technical program committees for several international conferences. He also served as a co-organizer of several international conferences.



608


Efficient Hardware Implementations for Tripling Oriented Elliptic Curve Crypto-System Mohammad Alkhatib, Adel Al Salem Abstract – This paper proposes several hardware designs and implementations for Tripling Oriented elliptic curve crypto-system (ECC) over GF (p). Projective coordinates are used to apply ECC's computations to avoid the time-consuming inversion operation. In order to improve the performance even further, parallel hardware designs to implement ECC operations in parallel are proposed. This plays a crucial role in speeding up ECC operations, compared to the known serial design. This research also presents several ECC designs by varying parallelization level for elliptic curve's computations. The purpose of this process is to provide efficient design solutions that fit different security applications according to requirements and available resources for certain application. Moreover, the NAF algorithm is used to perform scalar multiplication operation benefiting from the ability of NAF algorithm to reduce the average number of point addition operations. The proposed designs are implemented using VHDL, validated with ModelSim, and then simulated using Xilinx tool with target FPGA. The 4-PM design using homogenous coordinates achieves the best performance level for Tripling Oriented curve. Such design is significant for security applications that need high-speed ECC. Jacobean coordinates, on the other hand, shows the best performance results when applied with less parallelization levels as well as the serial design. Copyright © 2014 Praise Worthy Prize S.r.l. - All rights reserved.

Keywords: Elliptic Curves Cryptosystem, Time-Consumption, Resources, FPGA Hardware Implementation, Projective Coordinates, Security Applications

In elliptic curve cryptography, the original text is mapped into a point on the curve, and then ECC performs its operations on that point in order to yield a new point lying also on the curve. The new point represents the encrypted text. The basic operation in ECC is scalar multiplication operation, which consists of two major operations: point addition; which adds two different elliptic curve points, and point doubling; which adds an elliptic curve's point to itself [1], [4]. Cryptographic operations must be fast and accurate. In order to achieve this, elliptic curves cryptography uses finite field (also called Galois Field (GF) arithmetic. Two main finite fields are widely used in elliptic curves cryptography: Prime Field GF (p) and Binary Field GF (2n). This research focused on the GF (p). The main arithmetic operations in GF (p) are: modular addition, modular subtraction, modular multiplication, and modular division (inversion operation). The latter one requires finding the multiplicative inverse, which has been known to be the most costly and time-consumption operation in elliptic curves cryptography [4]-[6]. The use of projective coordinates systems to apply ECC operations instead of usual form of affine coordinates has been proposed as an efficient solution for inversion problem due to the ability of projective coordinates to eliminate the inversion operation by converting it to a number of multiplication

Nomenclature AT AT2 SM SA ECC PE PM PA SU FPGA

Area × Time Area × Time2 Sequential Multiplication Sequential Addition Elliptic Curve Crypto-system Parallelization Enhancement Parallel Multiplier Parallel Adder System Utilization Field Programmable Gate Array

I.

Introduction

In 1985, Koblitz and Miller introduced elliptic Curve Crypto-system (ECC) algorithm as an efficient alternative for RSA Algorithm. Since that time, ECC gained increasing interest by researchers in the field of cryptography and its applications. The security level for ECC depends on the difficulty of the known discrete logarithm problem for elliptic curves. ECC is motivated by the fact that it offers equivalent security level to those provided by conventional public key crypto-systems such as RSA using much smaller key sizes [1]-[3].



609

Mohammad Alkhatib, Adel Al Salem

operations. Three main projective coordinates systems are used in elliptic curves cryptography, which are Homogeneous, López-Dahab, and Jacobean coordinates systems [7], [8]. However, using projective coordinates to avoid inversion operation increases the number of multiplication operations needed to perform ECC computations. The multiplication operation is the second most time-consumption operation among elliptic curve operations. Hence, using serial (sequential) design implementation to perform ECC computations, although it consumes less area, increases the critical path delay of the crypto-system [4], [9]. Therefore, serial hardware designs are not suitable for security applications that require high-speed ECC such as multimedia applications [12]. Serial designs were proposed in [4], [10], [11], [18] and many others. The use of parallel designs, on the other hand, represents a great saving in time-consumption, compared to serial designs. The parallel designs utilize the inherent parallelism in ECC computations by performing these computations, specifically multiplications, in parallel in order to improve the performance (speed) of ECC. This improvement is achieved by reducing the number of sequential multiplication (SM) operations required to perform scalar multiplication. The time of sequential addition operations (SA) is negligible compared to the time of SM. However, the use of parallel hardware designs needs additional area and resources [12]-[14]. This work focuses on hardware implementations for elliptic curves, since they are faster and more secure than software implementations. The main components used to build the hardware designs of ECCs are multipliers and adders [13]-[15]. Authors in [16], [17] proposed a parallel hardware design for Standard (Weierstrass) ECC over GF (2n). In [16], a high-speed Elliptic Curve crypto-system architecture over GF (2n) has been proposed using two types of projective coordinates, which are homogenous, and Jacobean coordinates systems. Moreover, elliptic curve computations were mapped into parallel hardware design to be performed in parallel manner. The homogenous projection has shown the best performance results; the average number of multiplication cycles has been reduced in comparison to the other projections as well as the corresponding serial design. However, the presented designs consume more area and resources, and suffer low degree system utilization for the design using Jacobean projection. The AT cost is also high. As a result, presented hardware design is not suitable for implementations with limited resources. A Common aspect among researches presented in [7][12], [16], [17] is that the majority of them focused on the speed factor when designing ECC and neglected other factors. It is true that speed is an important factor but it is not the only one. The other factors such as system utilization, resources-consuming, area, AT2, and

AT costs in addition to performance factor play a crucial role in designing efficient ECC for different security applications. These factors have not been intensively investigated in the majority of previous researches. In [9], researchers emphasize that designers must insure that the hardware components specifically the multipliers will not remain idle when designing efficient ECC. This highlights the importance of considering system utilization. In [19] and [21], researchers have sown several designs for Standard ECC point addition and doubling operations respectively using different parallelization levels. several factors have been studied and improved. The highest performance level was achieved using homogenous coordinates with four parallel multipliers (PM) for point addition, and 5-PMs for point doubling, where the design using 2 PMs has shown the best degree system utilization for both operations. Almost all previous researches in the field of elliptic curve cryptography focused on the Standard (Weierstrass) elliptic curve and GF (2n), while there are many elliptic curve representations have not been widely considered in previous researches such as Hessian elliptic curve, Doubling Oriented elliptic curve, and Tripling Oriented Curve over GF (p), which is applied in this research. Moreover, relatively few studies have been conducted on the GF (p). Therefore, exploring such new curves is interesting and promising as it may lead to significant enhancements on several factors affecting the design and implementation of ECC for different security applications. Although the majority of researchers focused on studying the standard elliptic curve, a few researches that considered other forms of elliptic curves were also presented [18], [20], [22]-[24]. In [18], researchers have studied three elliptic curve representations over GF (p) which are: Doubling Oriented, Doche-Kohel-Icart Curves, and Jacobi-Quartics Curves. The curves have been implemented using three types of projective coordinates: homogeneous, López-Dahab, and Jacobean coordinates systems to eliminate the inversion operation. The proposing hardware architectures to implement elliptic curve point doubling operation has used PMs to accommodate the extra multiplication operations resulting from the use of projective coordinates and thus achieve higher performance level. Experimental results have shown that the proposed designs enhances the execution time for scalar multiplication operation compared to the corresponding designs using serial designs. However, authors have not considered other important factors such as the degree to which the system would be utilized, resources-consuming, and AT cost. In other words, the proposed designs are not suitable for applications where the speed (time-consuming) factor is not the first priority. While in [14], Edwards’ elliptic curve point doubling over GF (p) has been studied. The three projective coordinates mentioned in the previous research were also applied to Edwards curve to avoid the inversion operation.



610


Then, elliptic curve computations have been mapped into parallel hardware to speed up elliptic curve point doubling operations. However, These designs are considered costly choices, and the system utilization level is also low. It can be noted that the designs proposed in [14]-[18], [21] focus on point doubling operation only. Authors in [20] presented high-speed designs and architectures for the Binary Edwards ECC over GF (p) using projective coordinates. The homogenous coordinates have achieved the best performance level compared to other coordinates systems. However, in addition to consuming more area and resources, the proposed designs suffer low system utilization degrees. Moreover, the AT costs are high. Such designs seem also expensive for elliptic curves applications with limited resources. Although some previous researches have presented high-speed ECCs designs, these designs are not suitable for elliptic curve applications with limited resources and area. In some security applications, such as in the case of wireless sensors, the ECC is constrained by the available resources for a particular sensor, as well as the required performance level for that application. Thus, the use of five or four PMs in designing ECC, in despite of achieving high performance level, is not worthwhile for such applications. There is therefore a need to find several design solutions for ECC that satisfy different security applications based on the available resources and required performance level for a particular application. In the following section 2, projective coordinates are used to implement Tripling Oriented ECC operations, the computational schemas for proposed ECC designs are also shown. Section 3 presents synthesizing and implementation results for proposed designs. Conclusion and future work are presented in section 4.

where: = ( −

=

( )=

+ 3 ( + 1)

(3)

=

+ 2 ( + 1)] (2 )

3[

(4)

A. Homogeneous Projective Coordinates System In this projection, we substitute each (x,y) by (

(X/Z,Y/Z). The slope m in (4) = (

slope m=

)

. Let Q = 3[

+2

)

=

( + )]. Then the

, this had been used to find (X3, Y3, Z3) as

follows. Using the Eqs. (2) and (3): 3=

3= =

4

−

− 8 4

=

− 8 4 − ]−8 8

−

2

2

[12

−

=

In order to match the denominator for both x3 and y3 to the projection used we need to multiply X3 by , then we get the following results:

Y3= [12

In this section, elliptic curve computations had been studied and applied using three projective coordinates systems: homogeneous, López-Dahab, and Jacobean coordinates. The computational schemas for ECC point doubling had been also shown. Parallel hardware designs presented in this section uses different degrees of parallelism in order to provide a trade-off between performance (speed), resources-consumption, system (hardware) utilization, AT and AT2 factors. A complete description for these factors can be found in [20], [22]. It should be noted that the computational schemas for point addition operation have been presented in [19]. Let E be a Tripling Oriented elliptic curve over GF(p), then E can be defined by the equation: =

)−

The following subsections shows point doubling computations using projective coordinates systems.

Modeling and Equations

E:

(2)

where the slope:

X3= 2

II.

–2

[

− 8 −

]

]−8

Z3= 8 Inherited parallelism in homogeneous projection and elliptic curve's computations had been utilized the by using parallel hardware designs to apply ECC computations. Several design choices for Tripling Oriented ECC Point doubling operation were presented using different parallelization levels. We started by the design that accomplishes the highest performance level down to the serial design. The main hardware components used to build proposing designs are multipliers and adders, which are represented by circles and rectangles respectively in figures bellow. Each design had been also studied in terms of timeconsumption, resources, system utilization, AT, and AT2 factors. The computational schemas, that show how the data flows from inputs to outputs, for hardware designs using 4, 3, and 2 PMs are shown in Figures 1, 2 and 3 respectively.

(1)

Point doubling operation is computed as follows: P3 = P1 + P1 = 2P1 = (x3, y3)



611


Then the slope m= , this will be used to find (X3, Y3, Z3) as follows. Using the equations (2) and (3): = =

2

4

−

−

2

=

− 8 4

− 8 − 4 [12 − = 8

]−8

In order to match the denominator for both X3 and Y3 to the projection used we need to multiply Y3 by , then we get the following:

Y3= 2

X3= − 8 [12 − Z3= 4

] − 16

Fig. 1. The computational schema for Tripling Oriented ECC Point doubling with homogeneous coordinates using 4 PMs


B. López-Dahab Projective Coordinates System This projection substitutes each (x,y) by (X/Z,Y/Z2). In (4), the slope:

=

3

Let Q = 3[

+

+2

(

)

=

3[


Figs. 4, 5 and 6 show the computational schemas for Tripling Oriented ECC Point doubling with LópezDahab projection using 4, 3 and 2 PMs respectively.

+ 2 ( + )] 2

C. Jacobean Projective Coordinates System Here we substitute each (x,y) by (X/Z2, Y/Z3).

( + )].



612


The slope m =

=

.

Let Q =[ + 2 ( + )]. Then the slope m= , this will be used to find (X3, Y3, Z3) as follows. Using the Eqs. (2) and (3): =

=

3 2 =

9 4

−

−

2

9

3 [12

=

9

− 8 4

− 8 − 4 −9 ]−8

=

8

then X3, Y3 and Z3 can be calculated as follows: X3= 9 − 8 Y3= 3 [12 −9 ]−8 Z3= 2 Figs. 7, 8, and 9 show the computational schemas for Tripling Oriented ECC Point doubling with Jacobean projection using 4, 3 and 2 PMs respectively. A common observation about computational schemas presented above is that the computations in each level cannot commence unless the computations of previous level has been completed. The next subsection describes the hardware implementation and synthesizing for proposed designs.

Fig. 5. The computational schema for Tripling Oriented ECC Point doubling with López-Dahab coordinates using 3 PMs





613


Fig. 9. The computational schema for Tripling Oriented ECC Point doubling with Jacobean coordinates using 2 PMs Fig. 7. The computational schema for Tripling Oriented ECC Point doubling with Jacobean coordinates using 4 PMs

III. Results and Analysis

Fig. 8. The computational schema for Tripling Oriented ECC Point doubling with Jacobean coordinates using 3 PMs

This section presents and discusses results and outcomes of this research. Presented Tripling Oriented ECC designs are compared in terms of different factors. Table I shows a comparison between different designs with each projection. It can be noted from the table that a number of abbreviations are used such as: IC, SU, and PE, which stand for Idle Components, System Utilization, and Parallelization Enhancements respectively. It can be seen from Table I that the proposed 4-PM design with Homogeneous projection achieved the best (shortest) time delay. It is worth remembering that the time delay is estimated by the average number of multiplication cycles for both point doubling and addition operations, which are performed using NAF algorithm [19], [24]. Using NAF algorithm, the average case scenario for critical path delay can be calculated as follows: (1/3) × A + D, where A and D are the numbers of sequential multiplication levels in point addition and point doubling respectively [19], [22]-[24]. Furthermore, the 4-PM design showed similar performance level to that obtained by 5-PM design with much better system utilization and cost results. The main reason behind the enhancements in system utilization is that almost all hardware components, specifically multipliers, are fully utilized. The 4-PM design could be an effective design solution for security applications that require high-performance ECC.The other designs with homogeneous projection provided significant trade-off between area and time-consumption. The 2-PM design,



614


although needed less resources than designs with higher degrees of parallelism, it accomplished about 50% enhancements on performance with better AT2 cost, compared to corresponding serial design. As a consequence, this design can be considered for applications with limited resources. The 5-PM design, when applied with López-Dahab projection, showed better time delay results, compared to Jacobean projection. The latter, on the other hand, showed the least critical path delay and the best AT and AT2 results when applied using the serial design.

TSM= (cycles/bit) × m ×clockperiod where m is the key size, which is assumed to be 256 bits in our calculations. For instance, the time required to compute one sequential multiplication operation using GF (p) Tripling Oriented ECC design with projection (X/Z, Y/Z) at key size (m) = 256 bits can be calculated by: TSM= 1 × 256 × 9.079 = 2324.224 nano s (n s) Parallel designs and implementations for Tripling Oriented ECC are proposed in this research. In such implementations, the multiplication operations in each level are implemented in parallel once outputs from the previous level have been received. Therefore, the time above corresponds to the time of one SM level regardless of the number of multiplication operations that are performed in that level. This motivates the use of proposed parallel implementations, since it leads to significant improvements on the performance level, compared to the corresponding serial design implementation. It can be noted from Figure 1 that proposed Tripling Oriented ECC design consumed 3 SMs for each point doubling and 4 SMs for each point addition [19]. Thus, the time required for one point addition (TADD) and one point doubling (TDBLE) operation can be calculated as follows:

A. Simulation and Synthesizing Environment The scalar multiplication operation is usually used to calculate elliptic curve's point: KP from P, where P is a point on an elliptic curve, and K is a random integer. It is worth mentioning here that the non-adjacent form (NAF) was used to perform scalar multiplication, since it is faster and efficient for GF (p). The NAF algorithm decreases the average number of point addition operations to N/3, where N is the number of bits in K [4], [24]. The previously presented Tripling Oriented ECC designs were described using VHDL, which is a popular hardware description language. The VHDL code for our proposed designs was simulated using ModelSim (Mentor Graphics) simulation tool to validate its functional correctness. In respect to the hardware components, we used the known carry save adder (CSA) to avoid carry propagation, as well as carry save multiplier (CAM).Eventually, the Xilinx tool with the target FPGA (Field Programmable Gate Array) chip family chosen to be virtex5 (XC5VLX30) was then used to synthesize proposed designs, and to obtain results in terms of performance and resources-consumption.

TADD= 4×TSM= 9296.896 n s TDBLE= 3×TSM= 6972.672 n s It is worth mentioning that there is one inversion operation needed at the end of the scalar multiplication operation, in order to convert the point back to the usual affine coordinates form. The time of one inversion (TINV) can be estimated by the time of three SMs [19]-[22], as follows: TINV= 3 × TSM= 6972.672 n s

B. Implementation Results The performance of ECC processor can be described as the time needed to perform scalar multiplication (TKP). The time required to calculate one sequential multiplication (TSM) operation can be computed as follows:

Projection System Homogeneous Projection

López-Dahab Projection

Jacobean Projection

Finally, the NAF algorithm was used to compute the total time for scalar multiplication (TKP).

TABLE I COMPARISON BETWEEN PROPOSED ECC DESIGNS IN TERMS OF DIFFERENT FACTORS ECC design Resources Time Delay IC SU Cost Factors AT AT2 5-PM 5 M, 2 A 4.3 3M, 2A 79 % 21.5 92.5 4-PM 4 M, 2 A 4.3 2A 96 % 17.2 74 3-PM 3 M, 2 A 7 3M, 2A 79 % 21 147 2-PM 2 M, 2 A 8.7 2A 96 % 17.4 151.4 Serial 1 M, 1 A 17.3 100 % 17.3 299.3 5-PM 5 M, 2 A 5.7 8M, 3A 60 % 28 156.8 4-PM 4 M, 2 A 6 4M, 3A 73 % 24 144 3-PM 3 M, 2 A 6.3 3A 92 % 18.9 119.1 2-PM 2 M, 2 A 9.3 3A 92 % 18.6 173 Serial 1 M, 1 A 18.7 100 % 18.7 349.7 5-PM 5 M, 2 A 5.7 11 M, 3 A 47 % 28.5 162.5 4-PM 4 M, 2 A 5.7 7 M, 3 A 57 % 22.8 130 3-PM 3 M, 2 A 6 3 M, 3 A 73 % 18 108 2-PM 2 M, 2 A 8 1 M, 3 A 84 % 16 128 Serial 1 M, 1 A 14.7 100 % 14.7 216


PE 400% 400% 240% 200% 400% 400% 240% 200% 225% 225% 225% 180% -


615


TABLE II PERFORMANCE RESULTS FOR TRIPLING ORIENTED ECC USING THREE PROJECTIONS Projective 5-PM 4-PM 3-PM 2-PM Serial Coordinates System Design Design Design Design Design Homogeneous 2.577 2.577 4.16 5.147 10.681 Coordinates López-Dahab 3.368 3.565 3.761 5.54 11.074 Coordinates Jacobean 3.319 3.319 3.512 4.679 8.571 Coordinates 12

Tripling Oriented Curve 10

KP Total Time

8 6 4 2 0

5-PM

4-PM

3-PM

2-PM

Serial Design

Homogeneous

2,57

2,57

4,16

5,14

10,68

López-Dahab

3,36

3,56

3,76

5,54

11,07

Jacobean

3,36

3,36

3,56

4,74

8,69

Fig. 10. Performance characteristics for several ECCs proposed using three projective coordinates systems

It should be remembered that the NAF algorithm assumes that point addition happens in one third of the number of bits in m [4], [19], [22]-[24]. Therefore, the TKP can be computed as follows:

interesting trade-off between area and speed. These designs can be considered for security applications with limited resources. The 3-PM and 2-PM designs showed higher performance level when applied with Jacobean coordinates. Moreover, this coordinates system achieved much better performance results when implemented with serial design.

TKP= ((0.33× 256 × TADD) + + (256 × TDBLE) + TINV) × 10-6 = 2.57 m s Table II presents a comparison between proposed designs in terms of TKP. Note that TKP is shown in m s. Fig. 10 shows the performance characteristics for several ECCs proposed using three projective coordinates systems. It has to be mentioned that performance represents the total time delay for scalar multiplication operation (TKP). It can be seen from Fig. 10 that the 4-PM design obtained the shortest time delay when applied with Homogeneous projection. This design is recommended for security applications that need a high performance ECC. The reason behind this improvement is that the 4-PM leaded to less SM cycles for both point doubling and addition operations. The 5PM design, on the other hand, is not recommended to be used with homogeneous coordinates, since it consumed extra resources with no benefits in terms of performance. In other words, the Tripling Oriented curve using Homogeneous projection is saturated at the 4-PM design. Tripling Oriented curve also showed similar performance characteristics when applied using Jacobean Coordinates. The other presented designs provide

IV.

Conclusion and Future Work

Elliptic Curves Crypto-system (ECC) has been involved in different security applications. This research presented several hardware designs and implementations for Tripling Oriented ECC over GF (p). Elliptic curve computations were applied using three projective coordinates systems, which are homogeneous, LópezDahab, and Jacobean coordinates. The aim of this process is to eliminate inversion operation by converting it to successive multiplications. Furthermore, parallel hardware designs were used to improve ECC's performance. The proposed designs utilize the inherited parallelism in ECC's computations. It worth mentioning that NAF algorithm was used to perform scalar multiplication operation, since it decreases the average number of addition operations, which contributed in shortening the critical path delay. Proposed ECC designs were implemented using VHDL, and then ModelSim was used to validate its functional correctness. Finally, ECC designs were synthesized using the Xilinx tool with the



616


target FPGA. We provided several ECC designs by varying the degree of parallelism for elliptic curve's computations. This provided an interesting trade-off between several factors affecting ECC, such as performance, resources-consumption, system utilization, AT, and AT2. Such trade-off helps designers to select the most suitable ECC design for a wider range of security applications according to the requirements and available resources for specific application. A comprehensive comparison between proposed designs was also presented. The 4-PM design using homogeneous coordinates accomplished the highest performance level. This design may be considered for security applications that require high-speed ECC, such as multimedia applications. Designs with less parallelization levels could be efficient for applications with limited resources. Exploring different types of hardware components and projective coordinates systems sounds significant research direction, since it may lead to improve performance level for ECC.

[16]

[17]

[18]

[19]

[20]

[21]

References [1] [2] [3] [4] [5]

[6]

[7] [8] [9]

[10]

[11]

[12]

[13]

[14]

[15]

Wade Trappe, Lawrence. c, Introduction to Cryptography with Coding Theory. Washington, Pearson Prentice Hall, 2002. N. Koblitz, “Elliptic curve cryptosystem”, Mathematics of Computation, Vol. 48, pp. 203-209, 1987. V. Miller, “Uses of elliptic curves in cryptography”, Lecture Notes in Computer Science, Vol. 218, pp. 417-426, 1986. Blake, Seroussi, and Smart. Elliptic Curves in Cryptography. Cambridge University Press: New York, 1999 Chung, Sim, and Lee, “Fast Implementation of Elliptic Curve Defined over GF(pm) on CalmRISC with MAC2424 Coprocessor,” Workshop on Cryptographic Hardware and Embedded Systems, CHES 2000, Massachusetts, August 2000. A. Menezes, E. Teske, A. Weng, “Weak fields for ECC,” in the Cryptographer’s Track at RSA Conference (CT-RSA), LNCS 2964, 2004, pp. 366–386 Tanja Lange, “A note on López-Dahab coordinates”, Faculty of Mathematics, Technical University of Denmark, 2006. David, Nigel, and Jacques, “Projective coordinates Leak”, Applied research and security center, France. GuerricMeurice de Dormale, Jean-Jacques Quisquater. Highspeed hardware implementations of Elliptic Curve Cryptography: A survey. Journal of Systems Architecture 53 (2007) 72-84, by Elsevier. A. Satoh, K. Takano, “A scalable dual-field elliptic curve cryptographic processor,” IEEE Transactions Computers 52 (4) (2003) 449–460. G. Orlando, and C. Paar, "A scalable GF (p) elliptic curve processor architecture for programmable hardware,” Cryptographic Hardware and Embedded Systems - CHES 2001, Paris, France, May 14-15, 2001 Adnan Gutub and Mohammad K. Ibrahim., “High Radix Parallel Architecture For GF(p) Elliptic Curve Processor”, IEEE Conference on Acoustics, Speech, and Signal Processing, ICASSP 2003, Pages: 625- 628, Hong Kong, April 6-10. Adnan Gutub. Efficient Utilization of Scalable Multipliers in Parallel to Compute GF (p) Elliptic Curve Cryptographic Operations. Kuwait Journal of Science & Engineering (KJSE) 2007, 34(2): 165-182. L. Tawalbeh and Q. Abu Al-Haija. Speeding up Elliptic Curve Cryptography Computations by Adopting Edwards Curves over GF (P). International Journal of Security (IJS) 2009, CSC Journals, Malaysia, Vol.3, Issue.4, IJS-19. Q. Abu Al-Haija and Mohammad Al-Khatib. Parallel Hardware Algorithms & Designs for Elliptic Curves Cryptography to Improve Point Operations Computations Using New Projective

[22]

[23]

[24]

Coordinates. Journal of Information Assurance and Security (JIAS) 2010, By Dynamic Publishers Inc., USA, Vol.4, No.1, Paper 6: (588-594). Adnan Abdul-Aziz Gutub and Mohammad K. Ibrahim, "High performance elliptic curve GF (2k) crypto-processor architecture for multimedia", IEEE International Conference on Multimedia & Expo, ICME 2003, pages 81- 84, Baltimore, Maryland, USA, July 6-9, 2003. Adnan Gutub, “High Speed Low Power GF (2k) Elliptic Curve Cryptography Processor Architecture”, IEEE 10th Annual Technical Exchange Meeting, KFUPM, Dhahran, Saudi Arabia, March 23-24, 2003. L. Tawalbeh and Q. Abu Al-Haija. Enhanced FPGA Implementations for Doubling Oriented and Jacobi-Quartics Elliptic Curves Cryptography. Journal of Information Assurance and Security (JIAS), Volume 6, pp. 167-175 Khatib, M., Al-Haija, Q.A., Jaafar, A., Hardware architectures & designs for projective Elliptic curves point addition operation using variable levels of parallelism, (2011) International Review on Computers and Software (IRECOS), 6 (2), pp. 227-236. Mohammad Al-Khatib, Q. Abu Al-Haija, and Ramlan Mahmud. Performance Evaluation of Projective Binary Edwards Elliptic Curve Computations with Parallel Architectures. Journal of Information Assurance and Security (JIAS) 2011, By Dynamic Publishers Inc., USA, Vol.6, No.1, Paper1: (001-009). Mohammad Al-khatib, AzmiJaafar, and Q. Sbu Al-Haija. Choices on Designing GF (p) Elliptic Curve Coprocessor Benefiting from Mapping Homogeneous Curves in Parallel Multiplications. International Journal on computer science and engineering 2011, Vol.3, No.2, Paper 2: (467-480). Mohammad Alkhatib, Azmi B. Jaafar, ZuriatiZukarnain, and Mohammad Rushdan. On the Design of Projective Binary Edwards Elliptic Curves over GF (p) Benefiting from Mapping Elliptic Curves Computations to Variable Degree of Parallel Design. International Journal on computer science and engineering 2011, Vol.3, No.4, Paper 44: (1697-1712). Alkhatib, M., Jaafar, A., Zukarnain, Z., Rushdan, M., Trade-off between area and speed for projective edwards elliptic curves crypto-system over GF (p) using parallel hardware designs and architectures, (2011) International Review on Computers and Software (IRECOS), 6 (4), pp. 615-625. Al-Khatib, M., Jaafar, A., Zukarnain, Z., Rushdan, M., Hardware designs and architectures for projective Montgomery ECC over GF (p) Benefiting from mapping elliptic curve computations to different degrees of parallelism, (2011) International Review on Computers and Software (IRECOS), 6 (6), pp. 1059-1070.

Authors’ information Imam University, Faculty of Computer and Information Sciences, Department of Computer Science. Mohammad Al-Khatib is an assistant Professor in faculty of Computer and Information Sciences. He received the bachelor degree in Computer Science from IRBID National University. His Master degree was obtained from Depaul University in the field of Information Systems. He also achieved PhD in Computer Science (Security in Computing) from University PUTRA Malaysia (UPM). His research interest includes: information security, cryptography, Elliptic Curve algorithm, and information retrieval. Adel Al Salem is Computer Science Department in Al-Imam University. He received the bachelor degree from the College of Computer Science and Information Systems, King Saud University, KSA in 1994. His Master and PhD degrees were achieved from Kent State University, Ohio, USA in 1998 and 2001 respectively. His research interests includes: information systems, semantic operating systems, and information retrieval.



617


MSG SEVIRI Image Segmentation Using a Method Based on Spectral, Temporal and Textural Features Mounir Sehad1, Soltane Ameur1, Jean Michel Brucker2, Mourad Lazri1 Abstract – A new scheme for the classification of rainfall areas in convective and stratiform rain using MSG/SEVIRI (Spinning Enhanced Visible and Infrared) data is developed in this paper. The technique is based on spectral, temporal and textural properties of clouds. The introduced classification method uses as spectral parameters: brightness temperature BT (TIR10.8), brightness temperature differences BTDs (TIR10.8-TIR12.1, TIR8.7 -TIR10.8,TWV6.2-TIR10.8) during daytime and nighttime, temperature differences (TIR3.9 -TIR10.8 and TIR3.9-TWV7.3) during nighttime, and reflectances (RVIS0.6, RNIR1.6) during daytime. The textural information is based on the grey level rank approach where each pixel of the brightness temperature in the 10.8µm channel image is represented by a code which takes into account the relations between the spatial positions and the grey level ranks of the neighborhood pixels. The textural parameter of each pixel correspond to the occurrence frequency vector of a 24 pixel code possibilities computed within a window analysis of 31×31 pixels. The temporal parameter (RCT10.8) is the rate of change of brightness temperature over two consecutive images. The developed daytime and nighttime rain area classification technique (RACT-DN) is based on two multilayer perceptron neural networks (MLP-D for daytime and MLP-N for nighttime) which relies on the correlation of satellite data with convective and stratiform rain. The two algorithms (MLP-D and MLP-N) are trained using as reference convective and stratiform classification data from ground meteorological radar over northern Algeria. It was found that the introduction of temporal and textural parameters improved the results of discrimination between convective and stratiform areas. Copyright © 2014 Praise Worthy Prize S.r.l. - All rights reserved.

Keywords: Classification, MSG Image, Radar, Artificial Neural Network, Convective and Stratiform Clouds

Nomenclature C(i,j) Z R ri M Ni d X Od td Ed

 w η T SY SN RY RN

I.

Pixel code of the pixel P(i,j) Reflectivity factor Rainfall intensity Rank vector component Texture Pattern Occurrence frequency of the pattern Mi Training example Input vector of the multilayer perceptron Observed unit output Target unit output Error between the observed and the target outputs Weights vector of a node Learning rate Brightness temperature Number of estimated raining pixels Number of estimated non-raining pixels Number of observed raining pixels Number of observed non-raining pixels

Introduction

Rainfall estimation based on meteorological satellite data is one of the most intensely studied topics by meteorologists and hydrologists. The challenge is to produce accurate, high resolution and large extent rainfall accumulation maps based on the satellite data, which are essential for a variety of hydrologic applications [1]-[48]. The high spectral resolution of the Spinning Enhanced Visible and Infrared Imager (SEVIRI) on board the Meteosat Second Generation (MSG) satellites, with eleven 3-km-resolution channels and one 1-km resolution visible channel, offers the possibility of extracting the microphysical and dynamic structure of precipitating clouds allowing for an enhanced discrimination between convective and stratiform rain areas, and thus contributing to the improvement of the satellite rainfall estimation [1]. In this context, several techniques have been developed for rainfall process separation as a part of a satellite-based rainfall retrieval scheme in the midlatitudes using multispectral satellite data [2]-[3]. In a more recent study, developed schemes classify



618

Mounir Sehad, Soltane Ameur, Jean Michel Brucker, Mourad Lazri

convective and stratiform precipitation areas based on the high infrared spectral resolution of the MSG–SEVIRI [4]-[5]. The spectral features due to their physical importance have proved effective and simple in cloud classification [6]. However, they also encounter some drawbacks due to the spectral similarities of certain cloud features. Moreover, cloud texture refers to the physical appearance of the cloud. Such as a convective cloud which has a lumpy irregular texture, whereas cirrostratus cloud has a smoother and sometimes fibrous appearance in both visible and infrared channels. In a satellite image, information about the cloud types can be derived and mapped spatially by using a texture analysis [7]-[8]. Therefore, several studies have shown that the performance of the cloud classification algorithms is improved when textural features are taken into account [6]-[9]-[10]. Considering the advantages and drawbacks of the analysis schemes presented by Ameur [11], the grey level rank based-approaches are the best way to account the textural features. In addition, the incorporation of the rate of change in brightness temperature T10.8 over time provides information on cloud stage development. The aim of this paper is to propose a new operational technique for rain area discrimination in the Mediterranean region on a 15 min basis for MSG/SEVIRI daytime and nighttime data. The developed scheme is based on spectral, temporal and textural parameters. It is calibrated by instantaneous meteorological radar data using multilayer perceptron neural networks (MLPs). Artificial neural networks are widely used in precipitation remote sensing [12]-[13][14], and comparing to other statistical classification methods, the MLP algorithm does not require any a priori knowledge of the statistical distribution of the data [15]. The MLP is the statistical tool chosen to define the correlations between satellite measurements and classes of ground precipitation as estimated by weather radars. In order to take into account the variation of the diurnal cycle of clouds, the dataset is divided into daytime and nighttime data.

II.

Indeed, it is influenced by both the subtropical climate and the climate of mid-latitude systems [17]-[18]. The spatial distribution of precipitation is characterized by a very marked North-South gradient and a very low EastWest gradient. The rainy season extends from October to March, with maximum rainfall occurring during November-December. In the north, the climate is Mediterranean transit, marked by seasonal oscillations. The average annual rainfall is estimated at about 600 mm. The minimum rainfall is recorded in the southern regions. It is about 50 mm while the maximum is observed in the Djurdjura massif located in Kabylia and the massif of Edough located a little farther East, where it exceeds 1500 mm. The study area in this work is located in the north of Algeria, on domain with a radius of 250 km (see Fig. 1). In Fig. 1, the circle shows the domain of radar which coincides with the study area.

Fig. 1. The study area and the position of the weather radar of Setif. The circle shows the radar domain with a radius of 250 km

For this study, MSG/SEVIRI data together with corresponding ground-based radar data are required. II.2.

The dataset used in this work provided by the SEVIRI radiometer of Meteosat-8 in different frequency bands, the dataset are collected from November 2006 to March 2007 and November 2009 to March 2010. The MSG is a spinning stabilized satellite that is positioned at an altitude of about 36,000 km above the equator at 3.4°W. The SEVIRI radiometer gives every 15 minutes 12 images in the 12 available channels. We selected the channels sensitive to optical and microphysical properties of clouds (optical thickness, droplet size, cloud phase) as well as to the temperature of cloud tops, and those located in the spectral absorption bands mainly affected by the water vapor. These channels correspond to bands: visible (VIS0.6), near infrared (NIR1.6), water vapor (WV6.2, WV7.3) and infrared (IR3.9, IR8.7, IR10.8 and IR12.0). The raw image (Level 1.5) has a size of 3712 × 3712 pixels in each channel [19]. This corresponds to a spatial resolution at the image center of about 3 km. Each pixel is coded on 10 bits.

Study Region and Datasets II.1.

SEVIRI and Radar Data

Study Region

The training and validation of the developed technique are performed using SEVIRI/MSG and ground meteorological radar data for northern Algeria (see Fig. 1). Algeria is located on the South shore of the Mediterranean region; it is bordered on the East by Tunisia and Libya, on the South by Niger and Mali, South-West by Mauritania and Western Sahara and West by Morocco. This region has a particular orographic structure and special characteristics of the sea-land coast. Due to these geographical properties, its climate has a very complex spatio-temporal feature [16].



619


All pixels are geolocalized on a common grid in geostationary projection. The sub-satellite point corresponds to the pixel position (1856, 1856) on the image. For our case, we have predefined an area in the image of the Earth's surface; it corresponds to our study region (see Fig. 1). The radar data are provided by the ground-based C band radar network of The National Office of Meteorology (ONM). The Setif radar is installed near to the town of Setif, at 36 ° 11 'N, 5 ° 25' E and 1700 m of altitude, is one of seven Algerian meteorological radars. This is a Radar AWSR 81C in C-band, its operational frequency is 5.6 GHz. The displacement in azimuth is between 0 to 360 degrees in continuous and the movement in inclination is of -1° to 90°. Its polarization is linear and horizontal. The effective domain of radar is a radius of 250km. Meteorological radar data are collected at a temporal resolution of fifteen minutes and a spatial resolution of 1km in a format of 512×512 pixels. Each pixel is coded on four bits. Thus, it consist of 15 classes representing different reflectivity intensities which are all together considered as raining in the comparison with collocated satellite pixels and one class representing no raining. The physical parameter of the radar is the reflectivity factor, referred to as Z and expressed in (mm6m-3). The conversion of reflectivity factor Z into rainfall intensity R(mm/h) is obtained using the equation (1) adapted to our Radar and can also be converted into dBZ:

Z  300  R1.5

III.1. Spectral Parameters Brightness temperature ΤIR10.8 is an indication of the vertical extent of the cloud because, in general, brightness temperature of the system depends on the cloud-top height [2]-[21]-[22]-[23]-[24]-[25]. The brightness temperature difference TIR10.8-TIR12.1 being a good indicator of the cloud optical thickness, is very effective in discriminating optically thick cumuliform clouds from optically thin cirrus clouds[21][26]-[27]. Optically thick cumulus type cloud shows the smaller TIR10.8-TIR12.1 due to their black-body characteristics, while optically thin cirrus cloud shows the larger TIR10.8-TIR12.1 due to the differential absorption characteristics of ice crystals between the two channels [28]. It is expected that optically thick and deep convective clouds are associated with rain [29]. Even though the split window technique is very effective in detecting and removing optically thin cirrus clouds with no precipitation, it sometimes incorrectly assigns optically thick clouds like cumulonimbus in place of optically thin clouds [30]. The temperature difference TWV6.2-TIR10.8 is effective in distinguishing between high-level and low-level/midlevel clouds [31]. The 6.2-μm channel is dominated by atmospheric water vapor absorption. Low-level clouds produce temperatures at the 6.2-μm channel lower than their actual cloud top temperatures due to the absorption from water vapor above them. In contrast, their cloud-top temperatures at the 10.8-μm window channel are representative of actual cloud-top temperature since the atmosphere is transparent to this wavelength. As a result, TWV6.2-TIR10.8 tends to be very negative in sign for lowlevel clouds. In contrast, upper level thick clouds (being above most of this vapor and having absorption similar for both wavelengths due to ice crystals) produce temperatures at the 6.2-μm channel close to their actual cloud-top temperatures. In this case, TWV6.2-TIR10.8 usually takes very small negative values. Semitransparent ice clouds, such as cirrus, constitute an exception to this rule since their differential transmission cause larger negative differences. Positive differences may occur when water vapor is present in the stratosphere above the cloud top, which is a sign of convective cloud tops [32]-[33] as opposed to mere cirrus clouds. The brightness temperature difference TIR8.7-TIR10.8 can be utilized to gain information about the cloud phase [2][25]-[34]-[35]. The imaginary (absorption) component of the index of refraction, which is a direct indicator of absorption/emission strength, differs for ice and water at these two wavelengths [36]-[37]. More specifically, the difference in water particle absorption is small between the two wavelengths, but very large for ice particles [31][38]. Radiative transfer simulations show that for ice clouds, TIR8.7-TIR10.8 tends to be positive in sign, whereas for low-level water clouds TIR8.7-TIR10.8 tends to be small negative [37]. This simple parameter is adequate for classifying the cloud phase as either “ice” or “water”. We can expect ice cloud phase to be more associated with rain.

(1)

The scan interval for both data sets is 15 minutes. For the spatial comparison the radar data with an original spatial resolution of 1 by 1 km were projected to the viewing geometry of SEVIRI with a spatial resolution of 4 by 5 km in the study area. Because of discrepancies between the SEVIRI data and radar data, due to differences in observation time, parallax errors and collocation errors [20], the comparison of these types of data may be hampered. To reduce the imbalances mentioned above and find a better correlation, we performed a repositioning to SEVIRI data to coincide spatially with radar data. We also applied a resampling to radar data in order to have the same resolution as resolution of satellite data. The resolution is 4×5 km in the study region and is assumed constant due to low overlapped area observed by both sensors. Therefore, each SEVIRI pixel is collocated with 4×5 radar pixels. The time lag between the radar and the satellite is about 3 min. This small time difference does not require synchronization between the two data types.

III. Methodology The developed rain classification method is based on spectral, temporal and textural parameters. These parameters are given as follows. Copyright © 2014 Praise Worthy Prize S.r.l. - All rights reserved


620


The effective droplet radius (re) and the optical thickness (τ) of clouds are directly related to rainfall probability of a cloud; it is function to. The effective particle radius (re) defined by the ratio between the third to the second power of the droplet spectrum is taken in place of the actual droplet spectrum. The cloud optical thickness (τ) defined by the integration of the extinction coefficient integrated over the cloud geometrical thickness is considered representatively for the cloud geometrical thickness. During day-time, the values of re and τ considered for a rainfall intensity differentiation can be retrieved on a pixel basis using a combination of two solar channels, namely the VIS0.6 and NIR1.6 channel of MSG [25]-[4][39]. High values of reflectance RVIS0.6 correspond to high optical depth of cloud and low values of reflectance RNIR1.6 indicate large particles in the cloud. This means that a large re and τ is obtained when high values of RVIS0.6 coincide with low values of RNIR1.6. It should be noted that the retrievals are limited to satellite and solar viewing zenith angles smaller than 72°. During night-time, combinations of brightness temperature differences TIR3.9-TIR10.8 and TIR3.9-TIR7.3 are used to infer implicit information about of re and τ [2][4]-[39]. Indeed, for thin clouds with small or large particles, respectively (small or medium re and τ), brightness temperature differences reach the highest values. Thick clouds with small particles (medium re and τ) lead to small values of brightness temperature differences. In contrast, large particles together with a high optical thickness (high re and τ) results medium values of brightness temperature differences. Therefore, a raining cloud indicates mean values of brightness temperature differences.

method which implies the computation of the statistical distribution of grey level ranks, takes into account the relations between the spatial positions and the grey level ranks of the pixels in their neighborhood. Therefore, these relations can be used to show the local structure of the grey level repartition and describe the texture of objects in the image. For the 4-neighborhood of the pixel P(i,j), the grey level vector is [P(i-1,j), P(i,j-1), P(i, j+1), P(i+1, j)] and the labels 0, 1, 2 and 3 according to the way used to scan the image (see Fig. 2(b)). These labels are classified following the increasing order of the grey levels and the apparition order of the pixel, the ordered labels form the rank vector [r0, r1, r2, r3] (see Fig. 2(c)). The pixel code of P(i,j) is given by the following equation:

C  i, j   r0  40  r1  41  r2  42  r3  43 Spatial position

t

2

P(i,j) 255

20 Grey level

50

3

(a) 4-Neighboring pixel

40

0

20 255

255

1

2

50

3

(b) Grey level vector

The temporal parameter is the rate of change of brightness temperature in the 10.8 µm channel over two consecutive images (RCT10.8), this parameter provides information on cloud stage development [40] and is defined as:

TIR10.8  t   TIR10.8  t  1

0

40

1

III.2. Temporal Parameter

RCT10.8  t  

(3)

1

0

3

2

(c) Rank vector

C (i,j)=1×40+0x41+3x42+2x43=177

(2) (d) Pixel code

where Δt is the time difference between two consecutive images (i.e., 15 min for MSG).

Figs. 2. Steps of construction of a rank vector for a 4-neighbohood

The different combinations of the four labels (0, 1, 2, 3) lead to 24 rank vectors and thus 24 code possibilities namely patterns (M1, M2,… ,M24) for each pixel in the image(see Table I). The textural features of each pixel P(i,j) in the image are represented by a vector with 24 components (N1, N2, …. ,N24) where each value (Ni) is the occurrence frequency of the pattern (Mi ) corresponding to the computed codes of the pixels within the window analysis of 31 × 31 pixels which is centered in pixel P(i,j).

III.3. Textural Parameters Several rain classification methods proved the power of the textural parameters to improve the classification skill [7]-[9]-[11]. In this study, the grey level ranks analysis scheme is applied to extract textural features from MSG satellite data in the thermal infrared channel at 10.8 µm [11]. The principle of the rank based-method is to replace the grey level of a pixel by an appropriate code characterizing the pixels neighborhood. This Copyright © 2014 Praise Worthy Prize S.r.l. - All rights reserved


621


TABLE I 24 RANK VECTORS, PIXEL CODES AND PIXEL PATTERNS FOR A 4-NEIGBORHOOD Rank vectors Pixel code Pixel codes [r0 r1 r2 r3 ] Patterns [3 2 1 0] 27 M1 [2 3 1 0] 30 M2 [3 1 2 0] 39 M3 [1 3 2 0] 45 M4 [2 1 3 0] 54 M5 [1 2 3 0] 57 M6 [3 2 0 1] 75 M7 [2 3 0 1] 78 M8 [3 0 2 1] 99 M9 [0 3 2 1] 108 M10 [2 0 3 1] 114 M11 [0 2 3 1] 120 M12 [3 1 0 2] 135 M13 [1 3 0 2] 141 M14 [3 0 1 2] 147 M15 [0 3 1 2] 156 M16 [1 0 3 2] 177 M17 [0 1 3 2] 180 M18 [2 1 0 3] 198 M19 [1 2 0 3] 201 M20 [2 0 1 3] 210 M21 [0 2 1 3] 216 M22 [1 0 2 3] 225 M23 [0 1 2 3] 228 M24

Then, the weights of the nodes are corrected by using the following equation:

   w  w  Ed  w

(5)

where η is the learning rate, this parameter typically ranges from 0.2 to 0.8. 2) Back-propagation algorithm Xd= input; td= target output; Od= observed unit output 1. Initialize all weights to small random numbers. Until satisfied, do:  For each training example, do 2. Input the training example to the network and compute the network output 3. For each output unit ( k):

δk  Ok 1  Ok  tk  Ok 

(6)

4. For each hidden unit (h):

δh  Oh 1  Oh 



wh,k  k

(7)

k  outputs

III.4. Developed Scheme

5. Update each network weight (wi,j):

The technique used to delineate rain areas in the MSG image is an artificial neural network multilayer perceptron (MLP).

(8)

wi, j   j xi

(9)

where:

1) Multilayer perceptron algorithm A multilayer perceptron is a feed forward artificial neural network model that maps sets of input data onto a set of appropriate outputs. An (MPL) consists of three or more layers (an input and an output layer with one or more hidden layers) of nonlinearly-activating nodes. Each node in one layer connects with a certain weight (wi) to every node in the following layer. Generally, the activation function used is the sigmoid (see Fig. 3).

3) Application of MLP for rain areas delineation Two MLPs are created for the classification of the rain areas in the MSG/SEVIRI data. The first one (MLP-D) is used during the daytime and the second one (MLP-N) is applied during the nighttime. The daytime MLP-D scheme uses as input spectral data (TIR10.8, TIR10.8-TIR12.1, TIR8.7-TIR10.8, TWV6.2-TIR10.8, RVIS0.6 and RNIR1.6 ), the rate of change in TIR10.8 over time (RCT10.8) as temporal parameter, and textural measures (N1, N2 ,..., N24 ), where Ni is the occurrence frequency of the code pixel pattern (Mi). The nighttime MLP-N algorithm uses as input spectral data (TIR10.8, TIR10.8-TIR12.1, TIR8.7-TIR10.8, TWV6.2TIR10.8, TIR3.9-TIR10.8, and TIR3.9-TWV7.3), (RCT10.8) as temporal parameter, and Ni=1to24 as textural features. The number of the hidden layer neurons was selected using the network growing method [42] for the training phase. Therefore, the number of 35 neurons for both MLP-D and MLP-N minimized the network’s error functions (RMSE) after 800 iterations (see Fig. 5). Therefore, each MLP contains 31 neurons in input layer, 35 neurons in the hidden layer and 3 output neurons representing the three classes for convective rain, stratiform rain and norain (see Fig. 4). In this study, MLP-D and MLP-N were trained using SEVIRI data set over north Algeria for 2109 precipitation scenes from November 2006 to March 2007.

Fig. 3. Sigmoid unit

MLP utilizes a supervised learning technique called back propagation algorithm [41]. For each training example (d) in (D) the error between the target value (td) and the value produced by the perceptron (Od) is given by the following relation:

1  2 Ed  w   td  Od  2

wi, j  wi, j  wi, j

(4)



622


Standard verification scores such as, Probability Of Detection (POD), Probability Of False Detection (POFD), False Alarm Ratio FAR, Frequency BIAS index (Bias), Critical Success Index (CSI), equitable threat score (ETS) are used to evaluate the developed scheme. The verification scores calculated from Table II correspond to the results of discriminating raining from non-raining clouds, and the scores computed from Table III correspond to the results of the convective/stratiform rain area classification. TABLE II CONTINGENCY TABLE FOR RAINING/NON-RAINING AREA DISCRIMINATION, (SY=SC+SS) AND SN : NUMBER OF ESTIMATED RAINING AND NON-RAINING PIXELS, RESPECTIVELY ; RY AND RN : NUMBER OF OBSERVED RAINING AND NON-RAINING PIXELS BY RADAR, RESPECTIVELY Satellite Radar Raining Non-Raining Total Raining SY RY SN RY RY Non-Raining SY RN SN RN RN Total SY SN TSR

X7= RCT10.8 ; (X8, X9 ….X31) = (N1, N2… N24) a) MLP-D: Daytime: (X1, X2, X3, X4, X5, X6) = (TIR10.8, TIR10.8-TIR12.1, TIR8.7-TIR10.8, TWV6.2-TIR10.8, RVIS0.6, RNIR1.6) b) MLP-N: Nighttime: (X1, X2, X3, X4, X5, X6) = (TIR10.8, TIR10.8-TIR12.1, TIR8.7-TIR10.8, TWV6.2-TIR10.8, TIR3.9-TIR10.8, TIR3.9-TWV7.3) Fig. 4. Structure of MLPs convective/ stratiform rain classification algorithms: a)MLP-D ; b) MLP-N

TABLE III CONTINGENCY TABLE FOR CONVECTIVE/STRATIFORM RAIN CLASSIFICATION, SC AND SS:NUMBER OF ESTIMATED CONVECTIVE AND STRATIFORM PIXELS, RESPECTIVELY, RC AND RS: NUMBER OF OBSERVED CONVECTIVE AND STRATIFORM PIXELS BY RADAR, RESPECTIVELY Satellite Radar Convective Stratiform Total Convective SC RC SS RC RC Stratiform SC RS SS RS RS Total SC SS TSR

IV.1. Statistical Analysis The verification scores [(Bias)Y, (POD)Y, (POFD)Y, (FAR)Y, (ETS)Y, (CSI)Y] for discriminating raining from non-raining clouds are given as follows :  The Bias (Bias)Y describe the ratio between the estimated and the observed rain events :

Fig. 5. Variation of the RMSE for different numbers off neurones (30, 35, 40) as a function of training iterations number of MLP-D and MLP-N

IV.

Rainfall Identification Results and Performance Evaluation

 Bias Y



SY RY  SY RN  SY RY  S N RY

S  Y   0  , optimal: 1 RY

Models are validated against independent rainy days during November 2010 to March 2011, not used for training the rain area delineation algorithms. The evaluation was performed by comparison with instantaneous ground-based radar data collocated with SEVIRI data. The observation scenes made by the radar and satellite at a rhythm of 15 minutes are 14580, most of which are non-raining situations. To evaluate the potential improvement by the developed Rain Area Classification Technique during Daytime and Nighttime (RACT-DN), the validation scenes were also classified by the Enhanced Convective Stratiform Technique (ECST) [43] which is similar to the Convective Stratiform Technique(CST) [44] but additionally includes the water vapor channel temperature for a more reliable deep convective/cirrus clouds discrimination [45].

(10)

 The Probability Of Detection (POD)Y gives the fraction of pixels that have been correctly identified as raining by the satellite technique according to the radar product:

 POD Y



SY RY  SY RY  S N RY



SY RY   0 1 , optimal : 1 RY

(11)

 The Probability Of False Detection (POFD)Y indicates the fraction of the pixels incorrectly identified as raining by the satellite algorithm:



623


 POFD Y



SY RN  SY RN  S N RN



SY RN   0 1 , optimal : 0 RN

1) Discriminating raining from non-raining clouds results The scores of discriminating raining from non-raining areas are represented in Table IV, the daytime bias of 97% and nighttime bias of 95% indicate that the RACTDN slightly underestimates the rain areas detected by the radar compared to the ECST bias (daytime: 83%; nighttime: 83%). Moreover, 82% of the radar observed raining pixels are identified by RACT-DN during daytime and 78% during nighttime which indicates a best performance compared to the POD (daytime: 65%; nighttime: 62%) for ECST. The improvement in POFD and FAR during daytime and nighttime is more notable. Indeed, the POFD indicates that a lower fraction of the observed non rain events were misclassified as rain events by RACT-DN (daytime: 1%; nighttime: 2%) than by ECST (daytime: 6%; nighttime: 7%). Furthermore, the FAR denotes that a lower fraction of the pixels were wrongly classified as rain by RACT-DN (daytime: 17% nighttime: 21%) than by the ECST (daytime: 29%; nighttime: 31%). The overall good performance of RACT-DN during daytime and nigttime indicated by the good range of the verification scores is further supported by the CSI (daytime:74%; nighttime:72%) and the ETS (daytime:24%; nighttime: 22% ) which outperform the result of the ECST (CSI daytime:58%, CSI nighttime: 56%, ETS daytime: 15%, and ETS nighttime : 12%).

(12)

 The False Alarm Ratio ( FAR)Y describe the fraction of the satellite pixels that have been wrongly classified as raining pixels:  SY RN   FAR Y  SY RY  SY RN (13) SY RN    0 1 , optimal : 0 SY  The Equitable Threat Score( ETS)Y indicate how well the classified pixels by the satellite technique correspond to chance SYRY random:

 ETS Y



SY RY  SY RYrandom SY RY  S N RY  SY RN  SY RYrandom

 1    1 , optimal : 1  3 

(14)

with:

SY RYrandom

 SY RY  S N RY      S R  SY RN  RY  SY  Y Y  TSR TSR

TABLE IV STANDARD VERIFICATION SCORES COMPUTED FOR RAINING/NONRAINING AREA DISCRIMINATION TECHNIQUES: RACT-DN AND ECST DURING DAYTIME AND NIGHT TIME. THE SCORES ARE BASED ON 1936 PRECIPITATION SCENES WITH 8022506 RAINING PIXELS OBSERVED BY RADAR RACT-DN ECST Test Daytime Nighttime Daytime Nighttime (BIAS)Y 0.97 0.95 0.83 0.83 (POD)Y 0.82 0.78 0.65 0.62 (POFD)Y 0.01 0.02 0.06 0.07 (FAR)Y 0.17 0.21 0.29 0.31 (CSI)Y 0.74 0.72 0.58 0.56 (ETS)Y 0.24 0.22 0.15 0.12

(15)

 The Critical Success Index (CSI)Y which enclose all pixels that have been identified as raining by either the radar or the satellite technique:

SY RY SY RY  SY RY  S N RY  SY RN RY  SY RN (16)   0 1 , optimal : 1

 CSI Y



2) Convective/Stratiform rain area classification results: By analysing the validations scores of the classification of the rainy areas into convective and stratiform regions represented in Table V, the developed scheme (RACT-DN) performs better than ECST by exhibiting during daytime and nighttime higher POD, ETS, and CSI values as well as lower false alarms scores( FAR and POFD). More detailed, 82% (daytime POD) and 79% (nighttime POD) of the convective rain occurrences are identified by RACT-DN, while 29% (daytime FAR) and 30% (nighttime FAR) of the estimated events are wrongly classified as convective. Moreover, 16% (daytime POFD) and 18% (Nighttime POFD) of the observed stratiform events are misclassified as convective rain cases. ECST indicates lower value of POD (daytime: 62%, nighttime: 61%), and higher scores of FAR and POFD (daytime: FAR= 46%, POFD= 24%, and nighttime: FAR=46%, POFD=25%) than RACT-DN.

The verification scores [(Bias)C, (POD)C, (POFD)C, (FAR)C, (ETS)C, (CSI)C] corresponding to the convective/stratiform rain area classification are computed from the similar equations described above by replacing SY with SC, RY with RC, SN with SS and RN with RS. It should be noted that the number of the estimated and observed raining pixels are given by SY=SC+SS and RY=RC+RS, respectively. IV.2. Results and Discussions The verification scores computed for the 1936 daytime and nighttime validation scenes are summarized in 2 tables: Table IV for discriminating raining from nonraining clouds, and Table V for the classification of the precipitation areas into convective and stratiform regions.



624


TABLE V STANDARD VERIFICATION SCORES COMPUTED FOR CONVECTIVE/STRATIFORM AREA CLASSIFICATION TECHNIQUES: RACTDN AND ECST DURING DAY TIME AND NIGHT TIME. THE SCORES ARE BASED ON 1936 PRECIPITATION SCENES WITH 8022506 RAINING PIXELS OF WHICH 1824550 HAVE BEEN IDENTIFIED AS CONVECTIVE AND 6197956 AS STRATIFORM BY RADAR RACT-DN ECST Test Daytime Nighttime Daytime Nighttime (BIAS)C 1.10 1.12 0.83 0.83 (POD)C 0.82 0.79 0.62 0.61 (POFD)C 0.16 0.18 0.24 0.25 (FAR)C 0.29 0.30 0.46 0.46 (CSI)C 0.61 0.59 0.53 0.53 (ETS)C 0.23 0.22 0.18 0.15

The outperformance of RACT-DN to classify the rainy areas into convective and stratiform regions during daytime and nighttime is supported by the good range of CSI and ETS values compared to the ECST scores. As expected, the additional of textural information in the developed scheme result of a better classification of the rainy areas into convective and stratiform regions compared to the ECST algorithm based only on spectral information. To gain a visual impression of the performance of the proposed scheme, we presented in Figs. 6 the delineated rain area and in Figs. 7 the results of the classification of stratiform and convective clouds. The classification is performed for a scene of 06 January 2011 (11:45 UTC). Fig. 6(a) shows the brightness temperature in the channel IR10.8. Fig. 6(b), Fig. 7(b) show areas classified together by RADAR and RACT-DN, and Fig. 6(c), Fig. 7(c) show the regions identified simultaneously by RADAR and ECST. The number of misclassified pixels is more important for the ECST technique compared to RACT-DN technique. This visual results support the statistics results obtained previously.

Figs. 7. Delineated and classified rain area for the scene from 06 January 2011 (11:45 UTC): (a) BT10.8 image, (b) rain area classified by RADAR and RACT-DN, (c) rain area classified by RADAR and ECST

V.

Conclusion

A new technique for rain area classification during daytime and nighttime (RACT-DN) based on spectral, temporal and textural features using multispectral optical satellite data of MSG-SEVIRI was proposed. The method allows a proper detection of convective and advective/stratiform precipitation. The technique considers as spectral information 6 parameters: 2 parameters characterizing the optical and microphysical cloud properties (VIS0.6 and, NIR1.6 channels for daytime or both channel differences T3.9T10.8 and T3.9-T7.3 during nighttime), and 4 other parameters give information about the cloud phase (T10.8, T8-7-T10.8, T10.8-T12.1, and TWV6.2-TIR10.8). The textural parameter is based on the grey level rank approach where each pixel is represented by the occurrence frequency vector of a 24 pixel code possibilities computed within a window analysis of 31 × 31 pixels. Moreover, the temporal feature gives the rate of change of brightness temperature over two consecutives images. The parameters are merged and incorporate into the developed rain classification algorithm by using two daytime and nightime multilayer perceptron MLP-D and MLP-N, respectively. The two MLP (MLP-D and MLP-N) were trained using SEVIRI data set over north Algeria for 2109 precipitation scenes from November 2006 to March 2007 and the models are validated against 1936 independent precipitation scenes during November 2009 to March 2010 which are not used for training the rain area delineation algorithms. The results of the developed scheme were compared with both corresponding ground based radar and ECST algorithm.

Figs. 6. Delineated rain area for the scene from 06 January 2011 (11:45 UTC): (a) BT10.8 image, (b) rain area delineated by RADAR and RACT-DN, (c) rain area detected by RADAR and ECST



625


2006, Montreal, Quebec, Canada, July 9-12, pp. 435-440, 2006. [12] T. Bellerby, M. Todd, D. Kniveton, C. Kidd, Rainfall estimation from a combination of TRMM precipitation radar and GOES multispectral satellite imagery through the use of an artificial neural network. J. Appl. Meteor., 39, pp. 2115–2128, 2000. [13] D.I.F. Grimes, E. Coppola, M. Verdecchia, G. Visconti, A neural network approach to real-time rainfall estimation for Africa using satellite data. J. Hydrometeor, 4, pp. 1119–1133, 2003. [14] F.J. Tapiador, C. Kidd, V. Levizzani, F.S. Marzano, A neural networks–based fusion technique to estimate halfhourly rainfall estimates at 0.18 resolution from satellite passive microwave and infrared data. J. Appl. Meteor., 43, pp. 576–594, 2004. [15] Y. Hong, K. Hsu, S. Sorooshian, X. Gao, Precipitation estimation from remotely sensed imagery using an artificial neural network cloud classification system. J Appl Meteor 43, pp.1834–1852, 2004. [16] Lionello P, Malanotte-Rizzoli P, Boscolo R ,Alpert P, Artale V, Li L, Luterbacher J, May W, Trigo R, Tsimplis M, Ulbrich U, Xoplaki E, The Mediterranean climate: an overview of the main characteristics and issues, in: Mediterranean Climate Variability. Elsevier B.V, pp. 1–26, 2006. [17] R.M. Trigo, E. Xoplaki, J. Lüterbacher, S.O. Krichak, P. Alpert, J. Jacobeit, J. Sàenz, J. Fernàndez, J.F. Gonzàlez-Rouco, Relations between variability in the Mediterranean region and mid-latitude variability, in: Lionello P, Malanotte-Rizzoli P, Boscolo R (Eds.), Mediterranean Climate Variability. Elsevier, Amsterdam, pp. 179–226, 2006. [18] P. Alpert, M. Baldi, R. Ilani, S. Krichak, C. Price, X. Rodό, H. Saaroni, B. Ziv, P. Kishcha, J. Barkan, A. Mariotti, E. Xoplaki, Relations between climate variability in the Mediterranean region and the tropics: ENSO, South Asian and African monsoons, hurricanes and Saharan dust, in: Mediterranean Climate Variability. Elsevier B.V, pp. 149–177, 2006. [19] EUMETSAT, Applications of Meteosat Second Generation Conversion from Counts to Radiances and from Radiances to Brightness Temperatures and Reflectance, 2004, http://oiswww.eumetsat.org/WEBOPS/msg_interpretation/index.h tml. [20] G.A. Vicente, J.C. Davenport, R.A. Scofield, The role of orographic and parallax corrections on real time high resolution satellite rainfall rate distribution. Int. J. Rem. Sens. 23(2, pp. 221– 230), 2002. [21] H. Feidas, Study of a mesoscale convective complex over the eastern Mediterranean basin with Meteosat data. Eumetsat Meteorological Satellite Conference, Oslo, Norway, 5-9 September, 2011. [22] H. Feidas, A. Giannakos, Classifying convective and stratiform rain using multispectral infrared Meteosat Second Generation satellite data, 2011.Theor.Appl.Climatol.doi: 10.1007/s00704011-0557-y. [23] H. Feidas, A. Giannakos, Identifying precipitating clouds in Greece using multispectral infrared Meteosat Second Generation satellite data, 2010.Theor.Appl.Climatol, doi:10.1007/s00704010- 0316-5. [24] H. Feidas, G. Kokolatos, A. Negri, M. Manyin, N. Chrysoulakis, Y. Kamarianakis, Validation of an infrared-based satellite algorithm to estimate accumulated rainfall over the Mediterranean basin. Theor. Appl. Climatol.,2008, http://dx.doi.org/10.1007/s00704-007-0360-y. [25] B. Thies, T. Nauss, J. Bendix, Discriminating raining from nonraining cloud areas at mid-latitudes using Meteosat Second Generation SEVIRI daytime data.Atmospheric Chemistry and Physics, 8, pp. 2341–2349, 2008, http://dx.doi.org/10.1029/22008JD010464. [26] T. Inoue, On the temperature and effective emissivity determination of semi-transparent cirrus clouds by bi-spectral measurements in the 10-mm window region. J. Meteorol. Soc. Japan 63, pp. 88–99, 1985. [27] T. Inoue, A cloud type classification with NOAA-7 split-window measurements. J. Geophys. Res. 92, pp. 3991–4000, 1987. [28] T.Inoue, X. Wu, K. Bessho, Life cycle of convective activity in terms of cloud type observed by split window. In: 11th Conference on Satellite Meteorology and Oceanography, Madison, WI, USA, 2001.

The validation scores have shown the outperformance of the developed method during the daytime and nighttime for the discriminating of the raining from nonraining clouds, as well for the classification of the rainy areas into convective and stratiform regions. One of the main advantages of the developed method is the best performance of the rain area classification algorithms (MLP-D and MLP-N) during daytime and nighttime. Indeed, during night time the visible channel information (VIS0.6) and the near-infrared channel (NIR1.6) are replaced by the infrared parameters to gain the optical and microphysical cloud properties. In general, the results of this study showed that the combined use of spectral, textural, and temporal features in the MSG-SEVIRI can be beneficial for the classification of convective and stratiform precipitating clouds. A potential application of a new rainfall retrieval technique based on MSG/SEVIRI data is the improved rainfall detection in a high spatial and temporal resolution during daytime and nighttime. Therefore, RACT-DN can be beneficial for the tropical countries which have a lack of gauge networks or radars. The classification ANN-based system can be improved by using multisensory data such as MSG/SEVIRI and TRMM (Tropical Rainfall Measuring Mission).

References [1]

E.E. Ebert, M.J. Manton, Performance of satellite rainfall estimation algorithms during TOGA COARE. J Atmos Sci 55, pp.1538–1557, 1998. [2] B. Thies, T. Nauss, J. Bendix, Discriminating raining from nonraining cloud areas at mid-latitudes using Meteosat Second Generation SEVIRI nighttime data. Meteorological Applications, 15, pp. 219–230, 2008. [3] M. Lazri, S. Ameur, Y. Mohia, Instantaneous rainfall estimation using neural network from multispectral observations of SEVIRI radiometer and its application in estimation of daily and monthly rainfall. Advances in Space Research, 53(1), pp. 138-155, 2014. [4] M. Lazri, Z. Ameur, S. Ameur, Y. Mohia, J.M. Brucker, J. Testud, Rainfall estimation over a Mediterranean region using a method based on various spectral parameters of SEVIRIMSG, .Advances in Space Research, 52, pp. 1450-1466, 2013, http://dx.doi.org/10.1016/j.asr.2013.07.036. [5] H. Feidas, A. Giannakos, Classifying convective and stratiform rain using multispectral infrared Meteosat Second Generation satellite data. Theor Appl Climatol 108(3):613–630, 2012. [6] R. Kaur, A. Ganju, Cloud classification in NOAA AVHRR imageries using spectral and textural features. J Indian Soc Remote Sens 36, pp. 167–174, 2008. [7] Z. Ameur, S. Ameur, A. Adane, H. Sauvageot, K. Bara, Cloud classification using the textural features of Meteosat images. Int J Remote Sens 25, pp. 4491–4503, 2004. [8] M.JJ Uddstrom, W. Gray, Satellite cloud classification and rain rate estimation using multispectral radiances and measures of spatial texture. J Appl Meteor ,35, pp. 839–858, 1996. [9] A. Giannakos, H. Feidas, Classification of convective and stratiform rain based on the spectral and textural features of Meteosat Second Generation infrared data. Theor Appl Climatol 113, pp. 495–510, 2013. [10] Y. Shou, S. Li, S.Shou, Z. Zhao, Application of a cloud-texture analysis scheme to the cloud cluster structure recognition and rainfall estimation in a mesoscale rainstorm process. Adv Atmos Sc 23(5), pp. 767–774, 2006. [11] Z. Ameur, A. Adane, S. Ameur, Determination of the Grey Level Ranks for the Segmentation of Textured Images. IEEE ISIE



626


[29] T. Inoue, An instantaneous delineation of convective rainfall areas using split window data of NOAA-7 AVHRR. J. Meteorol. Soc. Japan 65, pp. 469–481, 1987. [30] T. Inoue, Day-to-night cloudiness change of cloud types inferred from split window measurements aboard NOAA polar-orbiting satellites. J Meteor Soc Japan 75, pp. 59–66, 1997. [31] H.J. Lutz, T. Inoue, J. Schmetz, NOTES AND CORRESPONDENCE Comparison of a split-window and a multi-spectral cloud classification for MODIS observations. J Meteor Society of Japan 81(3), pp. 623–631, 2003. [32] S. Fritz, I. Laszlo, Detection of water vapor in the stratosphere over very high clouds in the tropics. J Geophys Res 98(D12), pp. 22959–22967, 1993. [33] J. Schmetz, S.A. Tjemkes, M. Gube, L. van de Berg, Monitoring deep convection and convective overshooting with Meteosat. Advances in Space Research, 19(3), pp. 433–441, 1997. [34] S.A. Ackerman, K.I. Strabala, W.P. Menzel, R.A. Frey, C.C. Moeller, L.E. Gumley, Discriminating clear sky from clouds with MODIS. Journal of Geophysical Research, 103(D24), 1998,pp.32141–3215, . http://dx.doi.org/10.1029/1998JD200032. [35] K.I. Strabala, S.A. Ackerman, W.P. Menzel, Cloud properties inferred from 8–12 μm data. Journal of Applied Meteorology, 33, pp. 212–229, 1994. [36] B.A. Baum, P.F. Soulen, K.I. Strabala, M.D. King, S.A. Ackerman, W.P. Menzel, P. Yang, Remote sensing of cloud properties using MODIS airborne simulator imagery during SUCCESS: 2. Cloud thermodynamic phase. J Geophys Res 105(11), 2000, pp. 781- 792. [37] B.A. Baum, S. Platnick, Introduction to MODIS cloud products. Earth Science Satellite Remote Sensing,, pp. 87–108, 2006. [38] M.J. Pavolonis, A.K. Heidinger, T.Uttal, Daytime global cloud typing from AVHRR and VIIRS: Algorithm description, validation, and comparisons. J. Appl. Meteor., Vol. 44, Issue 6, pp. 804–826, 2005. [39] M. Lazri, S. Ameur, J.M. Brucker, J. Testud, B. Hamadache, S. Hameg, F. Ouallouche, Y. Mohia, Identification of raining clouds using a method based on optical and microphysical cloud properties from Meteosat second generation daytime and nighttime data, Appl Water Sci, 2013. DOI 10.1007/s13201-0130079-0. [40] G.A. Vicente, R.S. Scofield, W.P. Menzel, The operational GOES infrared rainfall estimation technique. Bull. Am. Meteorol. Soc. 79, pp. 1883–1898, 1998. [41] Nawi, N.M., Rehman, M.Z., Ghazali, M.I., Noise-induced hearing loss prediction in malaysian industrial workers using gradient descent with adaptive momentum Algorithm, (2011) International Review on Computers and Software (IRECOS), 6 (5), pp. 740748. [42] B. Krose, P. van der Smagt, An introduction to neural networks. University of Amsterdam, 1996, pp. 44–45. [43] C. Reudenbach, G. Heinemann, E. Heuel, J. Bendix, M. Winiger, Investigation of summertime convective rainfall in Western Europe based on a synergy of remote sensing data and numerical models. Meteorol.Atmos.Phys, 76, pp. 23–41, 2001. [44] R. Adler, A.J. Negri, A satellite infrared technique to estimate tropical convective and stratiform rainfall. Journal of Applied Meteorology, 27, pp. 30–51, 1988. [45] S.A. Tjemkes, L. van de Berg, J. Schmetz, Warm water vapour pixels over high clouds as observed by Meteosat. Beiträge zur Physik der Atmosphäre, 70(1), pp.15–21, 1997. [46] M. Satyanarayana, G. S. N. Raju, Modern Radars and the Generation of Specified Antenna Radiation Patterns, (2011) International Journal on Communications Antenna and Propagation (IRECAP), 1 (2), pp. 165-169. [47] Sharareh Kiani, Amir Mansour Pezeshk, Hossein Pourghassem, Design and Simulation of Monopulse Patch Linear Array for Passive SAR Satellite Tracking, (2012) International Journal on Communications Antenna and Propagation (IRECAP), 2 (1), pp. 45-50. [48] Yılmaz Kalkan, On the Advantages of Frequency-Only MIMO Radar, (2013) International Journal on Communications Antenna and Propagation (IRECAP), 3 (3), pp. 163-168.

Authors’ information 1

Laboratory for analysis and modelling random phenomena (LAMPA), University of Tizi Ouzou, Tizi Ouzou, Algeria. Tel: +213554587577 Fax :+21326218283 2

School of engineers: EPMI, Paris, France, EPMI - 13 Boulevard de l'Hautil 95092, CERGY PONTOISE Cedex. Tel: +33 130756040 Fax: +33 1 30 75 60 41



627


An Effective Selection of DCT and DWT Coefficients for an Adaptive Medical Image Compression Technique Using Multiple Kernel FCM Ellappan V., R. Samson Ravindran Abstract – Over the last two decades, great developments have been made in image compression approaches driven by a growing demand for storage and transmission of visual information. However, a number of publications have demonstrated good results of the compressed domain approach to image analysis. Nevertheless, very little work has been carried out on ROI based compression. In this paper, we have proposed an adaptive approach to compress the image without lossless version using selection of DCT and DWT coefficients and MKFCM. Our proposed approach consists of three stages, (i) Region segmentation (ii) Image compression (iii) Image decompression. At first, the input medical image is segmented by ROI, Non-ROI and background using MKFCM. Subsequently, the ROI region compressed by DCT and SPIHT coding and NonROI region is compressed by DWT and Huffman coding. The non-relevant regions are directly converted to zero. From the compressed regions like as ROI, Non-ROI and background, finally we obtain compression ratio to evaluate the proposed approach. Then, in decompression stage, the original medical image is extracted using the devised procedure. We can also see that our proposed image compression approach have outperformed by having better compression ratio of 4.6446 when compared existing technique only achieved 4.4326. Copyright © 2014 Praise Worthy Prize S.r.l. - All rights reserved.

Keywords: Image Compression, Discrete Wavelet Transform, Discrete Cosine Transform, ROI, Background, SPIHT, Huffman

dependent on certain important characteristics of the original images which have to be safeguarded after the compression task has been completed [3]. An image compression system comprises an encoder that employs the redundancies to characterize the image data in a compressed way, whereas the decoder is utilized to reconstruct the original image from the compressed data [6]. In medical domain, certain medical images such as Magnetic Resonance Imaging (MRI) and Computed Tomography (CT) necessitate lossless compression as even a negligible loss can lead to unfavorable effects. Moreover, attainment of high compression prediction is one of the methods so as to estimate current data from previously identified data [7]. For example, in medical image compression applications, diagnosis is valuable only when compression techniques preserve the entire pertinent and significant image data required [10]. In the past two decades, the frequently employed image compression technique for medical images is JPEG which integrates DCT transform with Huffman coding. During the last decade, on account of the necessity for enhancing the visual quality in compressed medical images, the wavelets (DWT) has continuously come out with flying colors in regard to image compression [18]. The vital qualities of wavelet transforms like Multiresolution representation, energy compaction, blocking

Nomenclature DWT ROI BG DCT MKFCM PSNR EBCOT MRI CT CRICM HOP

Discrete Wavelet Transform Region of Interest Background Discrete Cosine transform Multi-Kernel Fuzzy C-means Clustering Peak signal to Noise Ratio Embedded Block Coding with Optimized Truncation Magnetic Resonance Imaging Computed Tomography Cross-point Region for lossless Image Compression on Multiple bit planes Hierarchical Oriented Prediction

I.

Introduction

Image compression is indispensable for many of the telematic applications and it plays a critical role to ensure excellent quality of service [1]-[25]. The fast growth in the range and application of electronic imaging justifies interest for systematic plan of an image compression system and for furnishing the image quality necessary in various applications [5], [22]-[25]. The real issue is that while high compression rates are preferred, the applicability of the reconstructed images is



628

Ellappan V., R. Samson Ravindran

artifacts and de-correlation, have enabled the discrete wavelet transform (DWT) emerge as one of the most imperative techniques for image compression during the period. Triggered by the immense success of wavelet in medical image compression, the familiar and the sophisticated image compression techniques for medical images are enlarged in accordance with integer wavelet transform with Embedded Block Coding with Optimized Truncation (EBCOT) [8], called JPEG 2000 standard [9]. For the superior conservation of imperative image features and attainment of high compression ratios [10], simple DWT transform and fuzzy c-means clustering [11, 12] are proficiently integrated. Of late, the hybrid model of image compression has become the cynosure of the researchers. Considering above issues, in this paper, we have proposed an efficient image compression technique to improve the compression ratio. At first, the input medical image is given to multiple kernel based fuzzy cmeans algorithm ROI identification and detection. Then, for the ROI region, discrete cosine transform is applied to select coefficients and the Spiht encoder is applied to bit stream with the aim of doing lossy compressive in DCT coefficients. For the Non-ROI region, the wavelet transform is applied to select the wavelet coefficients and these DWT coefficients are given to the Huffman encoding that converts the image components into bit stream. The nonrelevant regions are directly converted as zero. Based on this encoding, header file is generated to perform decoding operation correctly. Then, decompression stage, the original medical image is extracted using the devised procedure. The rest of this paper is organized as follows: Section 2 gives a brief description of the literature survey. Section 3 describes the two efficient algorithms taken for proposed approach. Motivation of the proposed approach is explained in section 4. Section 5 explains the proposed image compression approach. Result and discussion is discussed in section 6. Conclusion is summed in section 7.

II.

Region for lossless Image Compression on Multiple bit planes (CRICM). It was a procedure of entropy coding having two significant segments such as modeling and coding. Now, with a view to reduce the inter pixel redundancy the cross-point region theory was made use of. Therefore they started using Jones’ technique to reduced coding redundancy. At last, it was found that in relation to parallel techniques the developed device furnished a meaningful improvement in compression ratio. They have used it mainly for cryptography as it was meant only for lossless compression. For resolution scalable lossless and near-lossless (NLS) compression, Jonathan Taquet and Claude Labit, [15] have deftly developed a hierarchical method. The adaptability of DPCM technique was integrated with new hierarchical oriented predictors which furnish resolution scalability with superb compression performances than the general hierarchical interpolation predictor or the wavelet transform. Anyhow, the planned hierarchical oriented prediction (HOP) was not actually competent on smooth images. They brought in predictors which were dynamically optimized by means of a least-square criterion. The investigational Lossless compression outcomes were achieved on a large-scale medical image database. The planned technique was significantly appropriate for NLS compression which presented an attractive rate– distortion tradeoff in relation to JPEG-LS and equivalent or a superior PSNR than J2K for a high bit rate on noisy medical images. Walaa M. Abd-Elhafiez and Wajeb Gharibi, [16], have admirably brought to light the efficiency of diverse block based discrete cosine transform (DCT). Now, before applying the process of DCT the color image was transformed in to YCbCr where Y is luminance component, Cb and Cr were chrominance components of the image. Moreover, in accordance with the categorization of image blocks to edge blocks and non-edge blocks the image modification was carried out. At this juncture, the edge block of the image was compressed with low compression and the non-edge blocks were compressed with high compression. The test outcomes exhibited the fact that the performance of the planned technique reached superb levels in view of the fact that the constructed images were marginally distorted and compressed with higher factor. Rongchang Zhao and YIDE Ma [17] have devised a technique and performed it on digital image segmentation. The planned approach was composed of neurons with spike coupling and gradient enhancement triggered by the knowledge of visual cortex. Moreover, it was identical to the one in the visual cortex which differentiates certain objects in real scene by capturing boundary information. By the formation of a fitting function it smoothes pixels within domain and enhances pixels at boundaries. The test outcomes of the innovative technique

Related Researchers: a Brief Review

Literature presents several methods for medical image compression. In this section, we attempt to analyze certain such works. Shaou-Gang Miaou et al. [13] have shrewdly structured a system that merges the JPEG-LS and an interframe coding with motion vectors. As the interframe correlation between the two adjacent images in a medical image series is generally not higher than in an usual video image sequence. Now, the interframe coding was initiated only when the interframe correlation is sufficiently high. With six capsule endoscope image sequences undergoing scrutiny, the planned technique was able to attain only average compression gains. T. T. Dang et al, [14], have deftly devised a technique for lossless encoding and decoding images, particularly medical images. The planned method was known as a Cross-point



629


vouchsafed the fact that it was effective contours and appropriate for region-oriented image compression. M. Beladgham et al, [18] have earnestly targeted at fine-tuning the quality of Medical images after the compression step. It was an important tool for storage or transmission of medical images. At this junction, for the purpose of compression they have employed curvelet and discovered that the transform furnished better result when compared to parallel comparison technique. The compression technique by curvelet was exploited by performing it on various kinds of medical images. At last, when assessed according to compression ratio and quality of the compressed image the planned technique furnished better result. Xiwen Zhao and Zhihai He [19]

have been able to keenly identify the key issue faced in image compression. The critical issues in image compression were edges, patterns, and textures. At this point, they gave shape to an efficient image compression technique in accordance with super-spatial prediction of structural units. It was also known as super-spatial prediction breaks this neighbourhood constraint which tried to locate an optimal prediction of structural segment within the entire image domain. Now, they took into account only the lossless image compression. Their extensive experimental outcomes demonstrate that the planned technique was very competitive and it outperformed the state-of-the-art image compression methods.

III. Two Algorithms Taken for Proposed Technique III.1. SPIHT Encoding and Decoding







1) Initialization: output n  log'2 max i, j  | ci, j |  ;   Set the LSP as an empty list, and add the coordinates  i, j   H to the LIP, and only those with descendants also to the LIS, as type A entries. 2) Sorting Pass: 2.1) for each entry (i,j) in the LIP do: 2.1.1) output Sn  i, j  ; 2.1.2) if Sn  i, j   1 then move  i, j  to the LSP and output the sign of ci, j 2.2) for each entry  i, j  in this LIS do: 2.2.1) if the entry is of type A then 

output Sn  D  i, j   ;



if Sn  D  i, j    1 then 



for each  k ,l   O  i, j  do: 

output Sn  k ,l  ;



if Sn  k ,l   1 then add  k ,l  to the LSP and output the sign of ck ,l ;



if Sn  k ,l   0 then add  k ,l  to the end of LIP;

if L  i, j   0 then move  i, j  to the end of the LIS, as an entry of type B,and go to Step 2.2.2); otherwise, remove entry  i, j  from the LIS;

2.2.2) if the entry is of type B then 

output Sn L  i, j  ;



if S n  L  i, j    1 then



add each  k ,l   O  i, j  to the end of the LIS as an entry of type A;



remove  i, j  from the LIS.

 3) Refinement Pass: for each entry  i, j  in the LSP, expect those included in the last sorting pass (i.e., with same n ), output the nth most significant bit of | ci, j | ; 4) Quantization-Step Update: decrement

n by 1 and go to Step 2.



630


III.2. Huffman Coding

archives zoom in dimension, medical image compression has become one of the vital segments to compress the data for specific channel bandwidths or storage requirements maintaining the acceptable quality. When analyzing the literature, so many techniques have been presented for lossy and lossless compression method with a need of better technique which is faster, memory efficient and simple surely suits the requirements of the user. Anyhow, lossy medical image compression is construed to be undesirable for carrying out diagnosis in many of the medical imaging applications, mainly because of quality degradation. Hence, with a view to enhance the diagnostic value of lossy compressed images, the region of interest (ROI) coding concept is launched in the projected application to enhance the quality in specified regions of interest only by performing lossless compression in these regions, maintaining the high compression in non-ROI of the image. In the case well-acknowledged lossless compression techniques, the compression ratio is approximately 80% of original size, and in the case of lossy encoders the compression ratio tends to be further higher (up to 530%) [20], [21] and still there are very many possibilities for considerable loss in data, which may hamper effective treatment, leading to loss of diagnostically vital segments of the medical image. In spite of a flood of methods intended for lossy and lossless compression, there is the necessity for universal acceptance in the image processing society. In line with this, V. K. Bairagi A. M. Sapkal [1] has proficiently put forward a method for a mechanical ROI-based, medical image compression. Now, ROI was determined by means of HIS transformation together with thresholding-based segmentation. Subsequently, ROI was compressed along with the lossless version of compression methods like arithmetic coding, whereas non-ROI is compressed by means of set partitioning in hierarchical trees (SPIHT) algorithm, employed after wavelet transform. Triggered by the work launched by V. K. Bairagi A. M. Sapkal [1], here, we have taken pains to formulate an image compression technique with the help of multiple kernel FCM [2], DCT transform and wavelet transform. At this juncture, multiple kernels FCM [2] is proposed for segmentation in place of thresholding-based segmentation technique to ROI detection as it is the most modern segmentation, which proved mettle by being efficient in its function.

Huffman coding is a capable source coding algorithm for source symbols that are not equally probable. In 1952, Huffman suggested a variable length encoding algorithm, based on the source symbol probabilities P  xi  ; where i  1, 2 , ,L . The algorithm is optimal if the prefix condition is met, because then the average number of bits needed to represent the source symbols is minimum. The steps present in the Huffman coding algorithm are as follows: i. Organize the source symbols in increasing order of their probabilities. ii. Bind the bottom two signals together and write the sum of the probabilities of the two symbols on the combined node. Label the two branches with “1” and “0” as shown in the Fig. 1.

Fig. 1. Huffman coding

iii. Consider this sum of probabilities as a new probability corresponding to a new symbol. Again form a new probability by binding together the two smallest probabilities. The total number of symbols is reduced by one each time two symbols are combined. The two branches of the two low probabilities bound together are always labeled as a ‘0’ and’1’. iv. Continue the procedure until only one probability remains (and it should be ‘1’ if the additions performed are correct). This completes the creation of the Huffman Tree. v. Follow the branches from the final node back to the symbol to identify the prefix codeword for any symbol. Read out the labels on the branches while the route is traced back. This gives the codeword for the symbol. By coding the symbols one at a time, the Huffman’s procedure generates the optimal code for a set of symbols and probabilities.

IV.

V.

Motivation of the Proposed Approach

Compression techniques are classified into two categories such as lossless and lossy methods. In the medical imaging field, lossy compression techniques, even though they turn out to the tune 10% compression ratio, are not normally found employed. The probable reason may be the potential damage of profitable clinical information affecting diagnosis. As medical imaging has metamorphosed into digital formats like DICOM and

Proposed Image Compression Technique

Image compression is a function of data compression that encodes the original image by means of a few bits. The underlying objective of image compression is targeted at bringing down the redundancy of the image and to store or transmit data in a competent form. Image compression coding is entrusted with the task of storing the image into bit-stream as compressed as possible and



631


to display the decoded image in the monitor as precise as possible. In fact, image compression in medical image is very complicated. We compress the image certain pixels are likely to be missing and hence attaining highest accuracy also very hard. Several techniques suggested in the literature have not met with success to the desirable level in view of the intricacy of medical images. Considering the above issues, in this paper, we have proposed a technique to compress the medical image with the lossless version. The overall diagram of the image compression is shown in Fig. 2. The proposed image compression technique consists of following three important sections to compress the medical images.  Segmentation of ROI, Non-ROI and Background using MKFCM.  Compression.  Decompression.

Fig. 3. Different part of the medical image

Fig. 4. Cross-sectional view of medical image

It is entrusted with the task of segmenting the image into segments in accordance with certain predetermined features of pixels in the image. The centre point of attraction of this document lies segmenting the image by means of multi-kernel fuzzy c-means clustering technique [2]. Normally, Fuzzy c-means (FCM) is a technique of clustering which enables a particular piece of data to be a member of multiple clusters. It tends to be considerably restricted to spherical clusters only. To solve this problem, kernel fuzzy c-means algorithm is used by mapping data with nonlinear relationships to appropriate feature spaces. Kernel combination, or selection, is essential task for effective kernel clustering. For most of the applications, it is not easy to find the right combination. In this paper a multiple kernel fuzzy c-means (MKFCM) algorithm which extends the fuzzy cmeans algorithm with a multiple kernel learning setting. By using multiple kernels and automatically adjusting the kernel weights, MKFCM is more significant to ineffective kernels and irrelevant features. Inevitably, this leads to the choice of kernels less crucial. To effectiveness of the proposed approach, here, two kernels such as quadratic and linear are used. Using this MKFCM algorithm, the original input medical image OI i  j  is clustered into three segments like as ROI

Fig. 2. Overall diagram of the proposed image compression technique

At first, the input MRI image is furnished to multiple kernel FCM algorithms for ROI identification and detection, as amply depicted by Fig. 2. Thereafter, in respect of the ROI domain, wavelet transform is performed to choose the wavelet coefficients and thereafter, provided to the SPIHT encoder for transforming the image segments into bit stream. The wavelet coefficients are selected for ROI detection as the visual quality is superior to that of DCT-based compression. As far as the non-ROI region is concerned, DWT transform is carried out to choose DCT coefficients and subsequently, Huffman coding is executed on bit stream with the target of performing loss-less compressive in DWT coefficients. The irrelevant domains are directly transformed as zero. In accordance with this encoding, header file is turned out to carry out decoding operation precisely. In the case of MRI or CT image, there are three significant segments such as ROI which represents the diagnostically vital segment, non-ROI image segment, and the background which comprises segments other than image contents which are very well illustrated in Fig. 3 with its cross-section well depicted in Fig. 4. V.1.

image

S ROI i, j  , non-ROI image

S NROI i  j  and

background image S BG i  j  . The general concept of MKFCM aims to minimize the same objective function as a single fixed KFCM. The objective function, cluster centers and membership functions for the proposed method are given below: N

Q

C

 uijm

com

 xi   ci

2

(1)

i 1 j 1

From the Eq. (1), the cluster centers and membership functions are derived and given by:

Segmentation of ROI, Non-ROI and Background Using MKFCM

n

 uij K H  x j ,ci  x j cj 

Segmentation has emerged one of the most vital parts in the overall image processing cycle and it constitutes a very fruitful and critical module in object segmentation.

j 1 n

(2)



 uij K H x j ,ci



j 1



632


observations: o More frequently occurred symbols will have shorter code words than symbol that occur less frequently. o The two symbols that occur least frequently will have the same length. The Huffman code is devised by merging the lowest probable symbols and this process is repeated until only two probabilities of two compound symbols are left and thus a code tree is generated and Huffman codes are obtained from labeling of the code tree. The detail explanation of Huffman coding procedure is illustrated in section 3.2. Finally we obtain the compressed Non-ROI image  non  ROI c .

1 / m 1

uij

1  K  x ,c    1  K  x ,c   H

i

j

c

H

j

(3) 1 / m 1

i

k 1

where:











K H x j ,ci  K1 x j ,ci  K 2 x j ,ci



  K 2  x j ,ci   Quadratic kernel K1 x j ,ci  Linear kernel

com  defined as a combination of multiple kernels  the d-dimension center of the cluster cj

Step 3: Computation of  ROI c  To compute

V.2.

Compression

Proposed image compression stage involves following important steps:  Input: Original Image OI i  j 

generate compressed ROC image  ROI c . Also, SPIHT encoder converts the image components into bit stream. SPIHT works by partitioning the wavelet decomposed image into significant and insignificant partitions based on the following function:

 output: compressed image CI i  j  Step 1: Segmentation using MKCM In the segmentation process, the region of the interest (ROI) is generated using Multiple K-FCM algorithm in order to improve the diagnostic value of loss-less compressed image. The input MRI image OI i  j  is

 

1 , if max i, j  ci, j  2n   Sn     elsewhere  0 ,

initially clustered into three segments such as ROI image S ROI i, j  , non-ROI image S NROI i  j  and background

(4)

where, 2n is the threshold, Sn   is the significant of a

image S BG i  j  using MKFCM algorithm.

set of co-ordinates,  and ci, j are the coefficient value

After ROI detection process, the major task is to compress the input image or medical image OI i  j 

at coordinates i, j . In this algorithm, three ordered lists are used to store the significance information during set partitioning. List of insignificant sets (LIS), list of insignificant pixels (LIP), and list of significant pixels (LSP) are those three lists. The detail explanation of SPIHT procedure is illustrated in section 3.1. Finally we obtain the compressed ROI image  ROI c from the encoding

with the lossless version. To obtain this result, we have individually compressed each segmented images as  ROI c ,  Non  ROI c and  BG c . Step 2: Computation of  Non  ROI c  To find  Non  ROI c , initially we have applied

process. Step 4: Computation of  BG c

decomposition to improve the compression process using discrete wavelet transform (DWT). The NonROI image S NROI i, j  is divided into four sub bands

Here, the remaining segmented regions, i.e. background regions except ROC and Non-ROC are selected as compressed background image  BG c .

HH, LL, HL and LH. From the four sub-bands, choose HL S ROI  i, j  sub-band to encoding process.  After the wavelet transform is applied to the non-ROI image S NROI i  j  , the Huffman encoding is utilized to

at first, DCT transform is

applied to select DCT coefficients and then, SPIHT is applied to bit stream with the aim of doing lossy compressive in DCT coefficients.  After the cosine transform is applied on ROI image S ROI i, j  , the SPIHT encoding algorithm is used to

 is the measured data,  is the any real number greater than 1.

N m

 ROI c ,

Step 5: Calculate compression ratio CI i  j  using

 ROI c ,  Non  ROI c and  BG c .

create

the compressed Non-ROI image Non  ROI  c . Generally the Huffman encoding is

From the above steps (2) (3) and (4), we have individually calculated the compression ratio using the Eq. (5). Finally, the compressed medical image or input image CI i  j  is obtained from the Eq. (6):

converted the image components into bit stream. Huffman code procedure is based on the two



633


Compression Ratio 

CI i, j  

VI.

Size of Original image (5) Size of compressed bit stream

 ROI c   Non  ROI c   BG c 3 V.3.

This section presents the results obtained from the experimentation and its detailed discussion about the results. The proposed approach of image compression is experimented with the medical image datasets and the result is evaluated with the compression ratio, PSNR, average difference, cross correlation, normalize absolute error.

(6)

Decompression

Proposed image de-compression following significant steps:  Input: compressed image:

stage

involves

VI.1. Experimental Setup and Evaluation Metrics The proposed image compression technique is performed in a windows machine having configurations Intel (R) Core i5 processor, 3.20 GHz, 4 GB RAM, and the operation system platform is Microsoft Wnidow7 Professional. We have used mat lab latest version (7.12) for this proposed technique.

 ROI c ,  Non  ROI c ,  BG c  Output: decompressed image DI i  j  Step 1: Calculation of  Non  ROI dc

 Non  ROI dc image  Non  ROI c , Huffman To compute

Result and Discussion

from the compressed

VII.

decoding and inverse

Evaluation Matrices

The formulas used to calculate the evaluation metrics PSNR and compression ratio are given as follows:  Compression ratio: Image compression ratio is defined as the ratio between the uncompressed size and compressed size as shown in Eq. (8):

DWT is utilized to generate the decompressed Non-ROI image  Non  ROI dc . Here reverse operation of Huffman coding and DWT is performed to obtain  Non  ROI dc . Step 2: Calculation of  ROI dc Here, the decompressed ROI image

Compression Ratio=

 ROI dc is

calculated from the compressed image  ROI c through

 PSNR: The description of PSNR is given in the following formulas:

the reverse operation of DCT and SPIHT decoder. Finally,  ROI dc is obtained using IDCT and SPIHT decoder. Step 3: Calculation of  BG c

PSNR  10 log10

The background regions except ROC and Non-ROC are selected as decompressed background image  BG dc . operation To calculate decompressed image Di i, j  , we merge the three decompressed regions like as step 1, 2 and 3. Finally, the resultant decompressed image Di i, j  is

 Wxy  Wxy* 

(9)

* Wxy  Compressed image pixel value at coordinate

 x, y 

obtained using OR operation as Eq. (5):

 ROI dc  Non  ROI dc

2 Emax  Ww  Wh

where: Ww and Wh  Width and height of the compressed image Wxy  Original image pixel value at coordinate  x, y 

Step 4: Calculate decompressed image Di i, j  using OR

Di i, j    BG dc

Size of Original image (8) Size of compressed bit stream

2 Emax  Largest energy of the image pixels (i.e.,

(7)

Emax =255 for 256 gray-level images)  Cross correlation: Cross correlation between original image OI i  j 

where:

Di i, j   Decompressed output image

and compressed image CI i  j  is:

 BG dc  Decompressed background region  ROI dc  Decompressed Region of Interest (ROI)  Non  ROI dc  Decompressed Non-ROI

M 1 N 1

OI i  j   CI i, j  

  OI  x  y  CI  x, y 

(10)

i 0 j 0



634


 Average difference: The average difference (AD) of the input and compressed image is given by:

AD 

1 MN

M

VII.2.

At first, the proposed approach segmented three regions such as Region of Interest (ROI), Non-ROI and Background. This section shows the experimental results of our proposed technique. Table I shows the experimental result of the image compression that contains the input medical image and corresponding segmented ROI, segmented non-ROI and background.

N

 OI  x  y   CI  x, y 

(11)

i 0 j 0

 Normalize absolute error: The normalized absolute error (NAE) is given by

VII.3. M

N

NAE 

M

(12)

N

 OI  x, y  i 1 j 1

VII.1.

Dataset Description

In this approach, three MRI images are used, which is publically available dataset. Here, three test images “size of 64×64” given in Figs. 5. Figs. 5 show the three input MRI images.

(i)

(ii)

Comparative Analysis

We have compared our proposed image compression scheme against existing technique [1]. The performance analysis has been made by formulated the tables of evaluation matrices such as compression ratio, PSNR, average difference, cross correlation, normalize absolute error. To robustness of the proposed compression technique, we have used evaluated through compression ratio, PSNR, average difference, cross correlation, normalize absolute error. The performance analysis of the proposed approach is evaluated and tabulated in Tables II, III and IV. In Table II, we have compared different wavelets against proposed and existing technique. Here, haar based proposed technique is better compression ratio when compared other wavelet based techniques. Table III illustrates the PSNR value against different wavelets against proposed and existing technique. From Table IV, we can see that our proposed image compression technique have outperformed by having better PSNR value of 42.99 dB and better crosscorrelation value of 0.7925 and better average difference value of 4.32 and normalized absolute error value of 0.139 and compression ratio value of 4.6446 when compared existing technique [1].

 OI  x, y   CI  x, y  i 1 j 1

Experimental Results

(iii)

Figs. 5. Dataset description of the proposed image compression technique

TABLE I SEGMENTATION RESULTS Input medical images

ROI

NON-ROI


Background


635


TABLE II COMPARATIVE ANALYSIS OF WAVELET VS. COMPRESSION RATIO Proposed Existing [1] Image Wavelet

Haar Daubechies 2 Daubechies 3 Daubechies 4 Daubechies 5 Coiflet 1 Coiflet 2 Coiflet 3 Coiflet 4 Coiflet 5

4.6446 4.3956 4.3687 4.3365 4.2992 4.3572 4.2698 4.1684 4.0650 3.9580

4.3994 4.1425 4.1425 4.1118 4.0799 4.1425 4.0485 3.9651 3.8535 3.7523

4.1067 3.9200 3.8910 3.8624 3.8328 3.8910 3.8037 3.7119 3.6205 3.5236

4.4326 4.6274 4.5537 4.5913 4.5724 4.6082 4.5537 4.434 4.434 4.3706

TABLE III COMPARATIVE ANALYSIS OF WAVELET VS. PSNR Proposed

4.2069 4.3840 4.3668 4.3497 4.3327 4.3668 4.3140 4.2589 4.2016 4.1407

3.9478 4.0916 4.0766 4.0600 4.0436 4.0766 4.0273 3.9745 3.9215 3.8654

Existing [1]

Image Wavelet

Haar Daubechies 2 Daubechies 3 Daubechies 4 Daubechies 5 Coiflet 1 Coiflet 2 Coiflet 3 Coiflet 4 Coiflet 5

Images

42.99 42.85 43.38 43.58 42.83 44.32 43.47 43.48 43.60 43.69

43.67 46.03 44.50 44.72 44.32 43.99 43.74 44.08 44.33 44.09

43.64 43.87 44.08 44.11 44.16 44.25 44.14 44.30 43.81 43.75

31.71 32.68 33.49 34.69 32.88 32.99 33.30 33.66 33.14 33.31

30.56 32.13 32.50 33.14 33.32 33.36 33.68 33.44 34.26 33.12

TABLE IV COMPARATIVE ANALYSIS OF DIFFERENT EVALUATION MEASURES Normalize absolute PSNR, db Cross-correlation Average difference error Existing Existing Existing Existing proposed proposed proposed proposed [1] [1] [1] [1]

31.84 32.76 32.82 32.44 32.26 32.51 34.04 34.49 34.50 34.78

Compression ratio proposed

Existing [1]

42.99

31.71

0.7925

0.7358

4.3291

-0.0370

0.1394

0.0463

4.6446

4.4326

43.67

30.56

0.7202

0.6929

4.3364

0.0349

0.1408

0.0511

4.3994

4.2069

43.64

31.84

0.7417

0.6685

4.3628

0.0051

0.1490

0.0485

4.1067

3.9478

compression process. The proposed approach is classified into three stages, such as region segmentation, compression and decompression. Initially, the input medical image is clustered into three regions like as ROI, Non-ROI and background utilizing MKFCM. Compression stage is generated using DCT, DWT,

VIII. Conclusion In this paper, an efficient technique is proposed to compress the image that saves a lot of bits in the image data transmission. The proposed technique is simple and effective method for image compression to improve the



636


[17] Rongchang Zhao and YIDE Ma,“A region segmentation method for region-oriented image compression”, Journal of Neurocomputing, Vol.85, pp.45-52, 2012. [18] M. Beladgham, I. Boucli Hacene, A. Taleb-Ahmed and M. Khélif, “MRI Images Compression Using Curvelets Transforms”, AIP Conference Proceedings, Vol. 1019, No. 1, pp.249, 2008. [19] Xiwen Zhao and Zhihai He, “Lossless image compression using super-spatial prediction of structural components”, In Proceedings of the 27th conference on Picture Coding Symposium, pp. 393396, Sue (3): 2013 289,2009. [20] Miaou, S.-G., Ke, F.-S., Chen, S.-C.: ‘A lossless compression method for medical image sequences using JPEG-LS and interframe coding’, IEEE Trans. Inf. Technol. Biomed., Vol.13, No.5, pp. 818-821,2009. [21] Baeza, I., Verdoy, A.: ‘ROI-based procedures for progressive transmission of digital images: a comparison’, J. Math. Comput. Model. Vol.50, pp. 849-859, 2009. [22] Sujatha, R., Ramakrishnan, M., Developing an effective and compressed hybrid signcryption technique utilizing huffman text coding procedure, (2013) International Review on Computers and Software (IRECOS), 8 (12), pp. 2940-2947. [23] Umaamaheshvari, A., Prabhakaran, K., Thanushkodi, K., Watermarking of medical images with optimized biogeography, (2013) International Review on Computers and Software (IRECOS), 8 (12), pp. 2974-2984. [24] Ghatasheh, N.A., Abu-Faraj, M.M., Faris, H., Dead sea water level and surface area monitoring using spatial data extraction from remote sensing images, (2013) International Review on Computers and Software (IRECOS), 8 (12), pp. 2892-2897. [25] Padmalal, S., Nelson Kennedy Babu, C., Automatic feature extraction using replica based approach in digital fundus images, (2013) International Review on Computers and Software (IRECOS), 8 (12), pp. 2917-2924.

Spiht encoding and Huffman encoding. Decompression stage is performed by IDCT, IDWT, Spiht decoding and Huffman decoding. From the results, we can infer that our proposed technique have achieved good results and is efficient.

References [1]

[2]

[3]

[4]

[5]

[6]

[7]

[8]

[9]

[10]

[11]

[12]

[13]

[14]

[15]

[16]

V.K. Bairagi A.M. Sapkal, "Automated region-based hybrid compression for digital imaging and communications in medicine magnetic resonance imaging images for telemedicine applications", IET Sci. Meas. Technol., Vol. 6, No. 4, pp. 247– 253, 2012. Long Chen, C. L. Philip Chen, and Mingzhu Lu, "A MultipleKernel Fuzzy C-Means Algorithm for Image Segmentation", IEEE transactions on systems, man, and cybernetics-part b: cybernetics, pp. 1263 - 1274, vol. 41, no. 5, 2011. Ansari and Anan, "Recent Trends in Image Compression and Its Application in Telemedicine and Teleconsultation", National Systems Conference, pp. 59-64, 2008. Sadashivappa and AnandaBabu, "Evaluation of Wavelet Filters for Image Compression", World Academy of Science, Engineering and Technology, Vol. 51, pp. 131-137, 2009. Sonja Grgic, Mislav Grgic and Branka Zovko-Cihlar, “Performance Analysis of Image Compression Using Wavelets", IEEE Transactions on Industrial Electronics, Vol. 48, No. 3, pp. 682-695, June 2001. Loganathan and .Kumaraswamy, "Medical Image Compression Using Biorthogonal Spline Wavelet with Different Decomposition", International Journal on Computer Science and Engineering Vol. 02, No. 09, pp. 3003-3006, 2010. Yao-Tien Chen and Din-Chang Tseng, “Wavelet-based medical Image compression with adaptive prediction”, In proceedings of International symposium on Intelligent Signal Processing and Communication Systems, Vol. 31, pp.1-8, 2007. Krishnan K.marcellin MW, Bilgin A, Nadar M, ”Prioritization of compressed data by tissue type using JPEG2000”, In proceedings of SPIE medical imaging, Vol. 5748, pp. 181-189, 2005. Ruchika, Mooninder Singh and Anant Raj Singh, “Compression of Medical Images Using Wavelet Transforms”, International Journal of Soft Computing and Engineering (IJSCE), Vol.2, No.2, pp. 2231-2307, 2012. D.A. Karras, S.A. Karkanis´ and D. E. Maroulis, “Efficient Image Compression of Medical Images Using the Wavelet Transform and Fuzzy c-means Clustering on Regions of Interest”, Proceedings of the 26th EUROMICRO Conference (EUROMICRO'00)-Vol. 2, pp.2469, 2000. Dimitrios A. Karras, “Efficient Medical Image Compression/Reconstruction Applying the Discrete Wavelet Transform on Texturally Clustered Regions”, International Workshop on Imaging Systems and Techniques, 2005. M.Tamilarasi and V. Palanisamy, “Medical Image Compression Using Fuzzy C-Means Based Contourlet Transform”, Journal of Computer Science, Vol. 7, No. 9, pp.1386-1392, 2011. Shaou-Gang Miaou, Fu-Sheng Ke, and Shu-Ching Chen, “A Lossless Compression Method for Medical Image Sequences Using JPEG-LS and Interframe Coding”, IEEE Transactions On Information Technology In Biomedicine, Vol. 13, No. 5, 2009. T.T. Dang, S.K. Nguyen, T.D. Vu and S. Higuchi, “Cross-point regions on multiple bit planes for lossless images compression”, IET Image Processing,Vol. 5, No. 5, pp. 466–471, 2011. Jonathan Taquet and Claude Labit, “Hierarchical Oriented Predictions for Resolution Scalable Lossless and Near-Lossless Compression of CT and MRI Biomedical Images”, IEEE Transactions on Image Processing, Vol. 21, No. 5, 2012. Walaa M. Abd-Elhafiez and Wajeb Gharibi, “Color Image Compression Algorithm Based on the DCT Blocks”, IJCSI International Journal of Computer Science Issues, Vol. 9, No. 4, No 3, 2012.

Authors’ information Ellappan V. (Venugopal) was born in Salem, Tamilnadu. He obtained his Bachelor’s degree in Electronics and Communication Engineering from Anna University Chennai. Then he obtained his Master’s degree in Applied electronics from Anna University of Technology Coimbatore, Tamilnadu, India, He has held the position of Assistant Professor and Researcher within the Centre for Electronics and Communication Engineering, Mahendra Engineering College, Tamilnadu, India. His main research interests are in the areas of image and video processing, and segmentation. He is life member Indian Society for Technical Education (MISTE), India. He is life member Intuition of Engineers India (IEI). Dr. R. Samson Ravindran was born in Sengam, Tamilnadu. He graduated in Electrical and Electronics Engineering (B.E,) from Anna University, Chennai at the degree level. Electronics and Control Engineering (M.S.,) as Masters from Birla Institute of Technology and Science (BITS) Pilani. Master’s in Business Administration (M.B.A) from the University of Madras. He is the Executive Director of the Mahendra Engineering Colleges is one of the most talented and a skillful administrator. He has been devotedly endeavored with a single mind to lift this institution to still higher levels of glory with a wide range of specializations at the different strata of study. He has been awarded Ph.D., for his Research work in the area of Solar Energy and his relentless pursuit of knowledge is exemplified in his endeavor as he had been awarded for his second Ph.D., in Bio-Engineering. He has visited countries like France, UK, Germany, Switzerland, Singapore, Malaysia, and Thailand to make study about higher Technical Education and Solar Power Projects. He has also presented many papers on Solar Power Projects, Bio-imaging and Telemedicine in various National and International Conferences and Journals.



637


Optimal Object Detection and Tracking Using Improved Particle Swarm Optimization (IPSO) P. Mukilan1, A. Wahi2 Abstract – In the current years, Object detection and tracking from the video sequence has turn out to be an exciting area of research. Using a single feature this research suggests new object exposure and tracking plan where video segmentation, characteristic extraction, feature clustering and object recognition are shared easily. The database video clips are partitioned into different shots, proceeding to implement the characteristic extraction. Feature extraction and tracking of the same video clips for the particular query clips are the two stages the organization is included. Mainly the contour of the frame can be examined by means of Enhanced Level Set algorithm. From the major contour detected the characteristics such as color, texture, edge density and motion are uttered. Obtained from parallel measures in the characteristic extraction, initially the motion feature is extorted by means of a competent motion estimation algorithm. As a result for tracking process here we employ Improved Particle Swarm Optimization (IPSO). For making sure high tracking presentation in the end, the tracked frames are collected by using a competent clustering algorithm. The expected strategy will be executed in MATLAB by means of a range of video clips and planned to be assessed. The appearance of the anticipated system will be considered by precision and recall measure. Copyright © 2014 Praise Worthy Prize S.r.l. - All rights reserved.

Keywords: Improved Particle Swarm Optimization (IPSO), Object Detection

Nomenclature hs  di 

Frequency of pixel values in bin di

Hs

Histogram of N bins

i i N PC , PR V B SW j

Mean

S FT

Area of the forward tracking

d1 , d 2 Vj(i+1) h j i 

Random values between [0,1] Velocity of particle j in the (i+1)th iteration Position of particle j in the i th iteration  Membership degree of datum d j to cluster



Prototype of cluster ki



 Distance between datum d j and prototype i

    ij  z

Introduction

Tracking can be described as the setback of estimating the trajectory of an object in the image plane as it goes around a scene. In other words, a tracker allocates reliable labels to the tracked objects in dissimilar frames of a video and furthermore depending on the tracking domain a tracker can too offer object centric information, such as orientation, area, or shape of an object [1]. Tracking is generally executed in the context of higher-level applications that need the location and/or shape of the object in every frame. Hypothesis is made to limit the tracking problem in the context of a specific application [2]. The tracking techniques are classified on the basis of the object and motion representations applied, offer detailed explanations of delegate techniques in each category, and observe their pros and cons. The employ of object tracking is relevant in the tasks of: motion-based recognition, automated surveillance, video indexing, human-computer interaction and traffic monitoring and vehicle navigation. Tracking objects can be difficult due to: loss of data caused by projection of 3D world on a 2D image [3], noise in images, complex object motion, non-rigid or articulated nature of objects, partial and complete object occlusions, complex object shapes, scene illumination changes and real-time processing necessities [4]. By impressive controls on the motion and/or

Standard deviation Number of sides in macro block, Pixels Object Background Area of the j th detection

ij   0,1 ki , i   t i ,d j

I.

   Set of all k clusters 1 ,2 ,..., k Fuzzy partition matrix Fuzzifier



638

P. Mukilan, A. Wahi

appearance of objects one can make simple tracking. For instance, almost all tracking algorithms guess that the object motion is soft with no sudden changes. Based on a priori information, one can additionally limit the object motion to be of steady velocity or steady acceleration. Earlier knowledge about the number and the size of objects, or the object appearance and shape, can furthermore be employed to make simpler the problem. The plan of an object tracker is to produce the trajectory of an object over time by finding its position in every frame of the video. At every time instant, Object Tracker [5] may moreover offer the entire region in the image that is engaged by the object. The assignment of identifying the object and launching correspondence among the object instances across frames can also be executed independently or together. In the initial case, using an object detection algorithm, feasible object areas in every frame are attained and then the tracker corresponds to the objects across frames [6]. In the last case, the object region and correspondence is equally estimated by iteratively revising object location and region information attained from earlier frames. Every tracking technique needs an object detection mechanism either in every frame or when the object first emerges in the video [7]. A general strategy for object detection is to employ information in a single frame. To decrease the number of fake detections, a few object detection techniques make use of the temporal information calculated from a sequence of frames [8]. This temporal information is generally in the form of frame differencing, which emphasizes changing regions in successive frames. Prearranged the object regions in the image, it is then the tracker’s task to execute object correspondence from one frame to the next to produce the tracks. In tracking, choosing the right characteristics takes part a significant role. In common, the most enviable property of a visual characteristic is its uniqueness so that the objects can be simply differentiated in the feature space [9]. Feature selection is strongly associated to the object representation. Generally all the tracking algorithms employ a mixture of these features: Color, Edges, optical Flow and Texture. Mostly characteristics are selected physically by the user depending on the application domain. On the other hand, the problem of automatic feature selection has obtained important attention in the pattern recognition community. The remaining of the paper is arranged as follows: Section 2 assesses some of the latest researches associated to our technique. Section 3 explains our suggested methodology that involves shot segmentation, feature extraction, and lastly object detection and tracking. Section 4 converses the results of the suggested technique. Section 5 is the final comments of the technique.

II.

approach to recognize pedestrians from laser range scans and use Joint Probabilistic Data Association particles filters to track moving pedestrians indoor. The related measurements were then strained out and classical scan registration and mapping methods in fixed environment were employed. In dynamic outdoor locations, Vu Trung et al. [11] suggested a real-time algorithm for simultaneous localization and local mapping with detection and tracking of moving objects from a moving vehicle equipped with a laser scanner, short-range radars and odometry. They launched a novel fast completion of incremental scan matching technique to correct the vehicle odometry that could function dependably in dynamic outdoor locations. A competent adapted directional lifting-based 9/7 discrete wavelet transform structure has been proposed by Jing-Ming and Chih Hsia [12] to further decrease the computational cost and protect the fine shape information in low resolution image. The experimental results documented that the suggested low-complexity MDLDWT plan could not offer more accurate detection rate for multiple moving objects, and the fine shape information could not be successfully protected for the real time video surveillance applications in both indoor and outdoor environments. A new background modeling has been suggested by Carlos Cuevas and Narciso Garcia [13] that is relevant to any spatio-temporal non-parametric moving object detection approach. Through a competent and vigorous technique to dynamically estimate the bandwidth of the kernels applied in the modeling, both the usability and the quality of earlier strategies are enhanced. Empirical studies on a large variety of video sequences showed that the suggested new background modeling decreases the quality of earlier strategies while preserving the computational necessities of the detection process. To find out the optimal threshold for the foregroundbackground segmentation and to learn background model for object detection, Simulated Annealing-Background Subtraction (SA-BS) has been suggested by Bahadir Karasulu and Serdar Korukoglu [14]. The attained performance results and statistical study demonstrated that the suggested technique is more preferable than regular BS technique. All the sampled points were accumulated in a single sampling map during the restarted SA-BS processes. After regarding all samples from all in-between states of the SA-BS process, not just the previous sample points to which it has converged, we loses the global convergence of SA-BS, and false peaks were identified in suggested SA-BS technique. Wang [15] improved the initial outdoor real-time system working out both Simultaneous Localization and Mapping (SLAM) and Detecting and Tracking moving Objects (DATMO) concurrently for urban environments from a ground vehicle. He employed an ICP-based

Review of Recent Researches

Paragios et al. [10] suggested a feature-based



639

P. Mukilan, A. Wahi

matching scan technique to correct the vehicle odometry and moving objects were identified based on a simple geometric study. He furthermore offered a mathematical framework integrating both SLAM and DATMO and demonstrated that they can be commonly beneficial from each other. Kirkpatrick et al. learned the Simulated Annealing, which springs from a property of solids where the heating was followed by the slow cooling and the solid reaches an almost perfect crystalline state giving up the minimum free energy [16]. The SA was created as the basis of an optimization method for combinatorial problems. When a solid was heated ahead of its melting point, its particles were positioned arbitrarily in the liquid phase. When it was cooled down, the particles reorganized themselves into lower energy configurations related with decreasing free energy. If the cooling was performed adequately slowly, at the zero temperature the solid would reach the global minimum energy configuration. A numerical hybrid local and global mode-seeking tracker has been suggested by Zhaozheng Yin and Robert T. Collins [17]. They regarded object detection as a global optimization problem and worked out it through Adaptive Simulated Annealing (ASA), a technique that evades becoming trapped at local modes and is much faster than complete search . To redetect the object and mend the local tracker, they used cluster study on the sampled parameter space. Their numerical hybrid local and global mode-seeking tracker was authenticated on challenging airborne videos with heavy occlusion and large camera motions. Schindler et al. [18] have offered an approach for multi-object tracking which regards object detection and space time trajectory estimation as a coupled optimization problem. It was devised in a hypothesis selection framework and erects upon a state-of-the-art pedestrian detector. At each time instant, it looks for the globally optimal set of space time trajectories which offered the best explanation for the current image and for all evidence gathered so far, while pleasing the constraints that no two objects may engage the similar physical space, nor make clear the similar image pixels at any point in time. Victorious trajectories hypotheses were fed back to direct object detection in future frames. The feature-based strategy was not able to labor in outdoor environment where different dynamic objects can never be explained by simple characteristics. The present researches could not offer more accurate detection rate for multiple moving objects, and the fine shape information could not be successfully protected for the real-time video surveillance applications in both indoor and outdoor environments. They decrease the quality of earlier approaches while sustaining the computational necessities of the detection process. The present multi-object tracking system which effected in pleasing the constraints that no two objects may engage the similar physical space, nor make clear

the similar image pixels at any point in time.

III. Proposed Methodology for Object Detection and Tracking III.1. Object Detection and Tracking In the video data study, the moving object detection and tracking is a significant process. The movement of the object from the video sequence in an appropriate dimension is essential for competent tracking and detection of the object. The moving object detection and tracking can be primarily made based on the background estimation. The camera vibration and calibration is a major subject in detection and tracking of the moving object from any video sequence. To work out all these problems we have assured a competent object tracking system with the help of optimization methods. III.2. Steps Involved in Our Proposed Method The suggested technique of motion based video object detection and tracking system contains the processes such as shot Segmentation, Feature Extraction and Tracking. The initial step in our strategy is to fragment the database video clips into dissimilar frames or shots. Then is the feature extraction step. In our suggested technique we extort different characteristics from the segmented image such as color feature, edge density feature, and texture feature and motion feature estimation. Using a competent motion estimation algorithm, the motion feature is extorted. Next, obtained from color quantization color feature is attained. Then for the objects presented in the database video clips edge density feature is attained, moreover the texture feature is abstracted by texton based method. By means of using together forward and backward tracking procedure, gained from these feature attained, the object will be scrutinized and the observed objects will be tracked. For estimated motion of an object from computed images in a video sequence visual tracking often encompasses an optimization process. Thus for tracking procedure here we employ Improved Particle Swarm Optimization (IPSO). The flow diagram for our suggested technique shown beneath. III.3. Feature Extraction When the key in information to an algorithm is too large to be processed and it is expected to be disgracefully redundant next the key in data will be changed into a reduced representation set of characteristics. Feature extraction is the change of input information into a set of features [19]. Extortion of image features and employ of these features to signify image visual content is usually termed as feature extraction. Feature extraction engages reducing the amount of



640

P. Mukilan, A. Wahi

resources necessary to explain a large set of data precisely.

Q

SD, i 

1  pik  i 2 Q k 1



(3)

The feature vectors of the values is erected as:



Fv  1 , 2 ,...,  K , 1 , 2 ,..., K



(4)

These feature vectors of the complete images are erected and accumulated in database. III.3.2. Texton Based Feature Extraction In the Fig. 2, Texton-based feature extraction is based on the texton frequency histogram. The erection of texton-based extraction contains three main stages like construction of a texton codebook, computation of a texton frequency histogram, and training of the classifier based on the texture frequency histograms. The common impression of the texton feature extraction is as illustrated beneath.

Fig. 1. Block diagram of our proposed method

III.3.1. Color Feature Extraction Using Color Quantization Using the color quantization technique, the color feature is extorted where histograms of meticulous images are being extracted. The histogram is described as the frequencies of the pixels in grayscale image. The quantization is a procedure in which the histogram is separated into levels or bins [20]. Computation cost for the feature extraction in these 256 levels will be high as grayscale image contains 256 levels. To decrease the computation cost, the histogram of image is diminished to dissimilar bins. The histogram is next quantized into N bins such that:

Fig. 2. General overview of Texton Feature Extraction

H s  hs  d1  ,hs  d 2  ,........,hs  d N 

III.3.2.1. Texton Codebook

(1)

A texton codebook is commonly a compilation of textons that can be employed to describe the texture images. Small texture patches in the texture images are extorted from random positions for codebook creation. To compose the pixel values of the texture patches invariant to lighting changes, they are standardized. Textons are attained by changing the texture patches to a suitable image representation. An entire texton codebook is erected by repeating this process for every texture class in the texture dataset. If the texture dataset is a correct subset of real-world textures, the texton codebook encloses the most significant textons that take place in real-world textures.

where hs  di  is the frequency of pixel values in bin di and H s is the histogram of N bins. The color features are employed for recovery of the same images. The data about the intensity level distribution of an image is offered by these color features. With the assist of intensity levels, the mean and the standard deviation can be prepared in the histogram bins. The mean and the standard deviation can be computed with the assist of the beneath expressions: Q

Mean, 1 

1 pik Q k 1



(2)



641

P. Mukilan, A. Wahi

III.3.2.2. Texton Frequency Histogram

beneath:

Frequency histogram is carried out after codebook construction. The texton measure is estimated using a texton frequency histogram that calculates the comparative frequency of textons from the codebook in a texture image. From a texture image, a texton frequency histogram is erected by scanning over the texture image and extorting small texture patches. The small texture patches are changed to the image representation that is applied in the codebook in order to attain a compilation of textons. In order to recognize the most similar texton from the codebook, each extracted texton is compared to the textons in the codebook and the texton frequency histogram bin related to this texton is incremented. After normalization, the texton frequency histogram structures a feature vector that models the texture.

Mean Square Error  MSE  =

1 N

2

N 1 N 1

   PC  PR 

2

(5)

p 0 q 0

Here N is the number of sides in macro block, PC and PR are the pixels that are being compared from the current block and the reference block.

III.3.3. Motion Feature Extraction for Motion Estimation in Video Normally motion estimation is the process of identifying the motion vectors which forms the transition from one frame to other in a video sequence. The object detection and tracking in any video can be identified based on the motion of the object at repeated sequences and this can assist us in identifying the object. We use block matching algorithm for motion estimation process in our suggested technique.

Fig. 3. Macro block with side 16 and search parameter P=7

III.4. Moving Object Detection and Tracking The moving objects are identified from each frame. There are dissimilar fundamental techniques used for moving object detection like:  Background Subtraction.  Temporal Differencing.  Optical Flow. The background subtraction technique is a simple technique for moving target detection. In background subtraction technique, it is assumed that the background is fixed so that the background does not vary with the number of frames. Initially the difference between the object V and the background B is computed using the formula:

III.3.3.1. Block Matching Algorithm The major contemplation behind the motion estimation is that the samples which correspond to objects and background in a frame of video sequence move inside the frame to form related objects on the successive frame. We separate the current frame into a matrix of ‘macro blocks in block matching process. These macro blocks are then compared with the same block and its nearby neighbors in the preceding frame to form a vector that state the movement of a macro block from one location to another in the preceding frame. Thus the motion estimated in the current frame can be acquired by computing the movements of all macro blocks in a frame. The search area in the macro block is usually reserved up to P pixels. This is known as the search parameter and this will be bigger for larger motion which in turn takes more execution time. Generally the macro block is a square of side 16 pixels and the search parameter P is 7 pixels. The general block matching idea is demonstrated in Fig. 3. The corresponding of this macro block is based on the cost function of the different macro blocks. The macro blocks whose output cost function is the least cost is the one that matches next to the present block. There is different cost function accessible and the least expensive cost function like Mean Square Error (MSE) is chosen for the cost calculation in our technique. Based on these MSE values we counterpart the blocks. The appearance for the Mean Square Error (MSE) is specified the eq.

D  m,n   V  m,n   B  m,n 

(6)

Now threshold the difference using the formula specified beneath:

1 , D  m,n   T T  m,n    0 ,otherwise

(7)

Using the gray histogram, the threshold can be selected by taking the bottom value between the two peaks as the threshold. Fundamentally Object tracking is applied to find the location of the target in dissimilar frames in a series of images. Choosing good target features and applying suitable searching techniques is the main work of target tracking. We use the Image Difference Algorithm for Moving Object Detection and Tracking in our work.



642

P. Mukilan, A. Wahi

III.4.1. Object Tracking

technique is the Improved Particle Swarm Optimization (IPSO). The IPSO is more competent than the normal PSO in terms of the fitness with the optimization efficiency. The IPSO mechanism is made clear in the beneath section.

To attain a good detection rate on each shot of a video frame, the detection and tracking were united and some rules are formed to obtain an entire tracking process. The tracking is of two kinds: 1) Forward Tracking. 2) Backward Tracking.

III.5. Tracking of Object with the Aid of Improved PSO PSO generated from the simulation of communal behavior of birds in a flock. In PSO, each particle flies in the search space with a velocity normalized by its own fast memory and its companion’s flying practice. Each particle has its purpose function value which is found out by a fitness function [21]. PSO is an evolutionary computation technique which is much associated to that of the Genetic Algorithm where a certain system is initialized by a population of random solutions. Randomized velocity is moreover assigned which compile a particle along with each potential solution in PSO. Each particle follows its coordinates in the problem space in link with the best solution. At this point the fitness value is also considered for the additional process. This fitness value is denoted to as pbest. The location of these solutions is regarded as gbest. In our proposed technique we have used an adapted version of PSO. In this improved PSO we have assigned worst case in addition along with the best case and also cross over operation is furthermore comprised after the fitness choice which would more increase the possibility of choosing the most outstanding particle.

III.4.1.1. Forward Tracking On each frame the forward tracking process is carried out, beginning from frames where the object have been identified. While tracking, similar object may be identified several times in a shot which can result in multiple tracking of the object, which may result in over time consumption. To overcome this problem, some tracking rules are applied to recognize whether the detected objects are multiplied or not. The rule is commonly based on the percentage of overlap between the identified object and the one from the forward tracking in the similar frame which is symbolized as follows:

S FT  max j

 FT W j 



min SW j ,S FT

(8)



where SW j is the area of the jth detection and S FT is the area of the forward tracking. In addition S

 FT W j 

represents the area recovered by the detection process. III.4.1.2. Backward Tracking Backward Tracking is executed on each frame to offer an additional set of object being tracked. The backward tracking is very constructive in case the object is not identified at the beginning but in the center of the frame. The forward tracking frequently signifies the object tracking from the detected frame to the end the whole shot while backward tracking presents the unnoticed result from the initial frame of the shot to the frame in which the last object detection has executed. Further the backward tracking can moreover proves to be successful when the forward tracking fails to set the position of the object in a particular frame. This may be owing to occlusion, bad illumination or due to tracker sticks to the background. That is when an object in a frame1 is not tracked properly and the similar object is tracked in frame 5, the data will be broadcasted back and will give tracking of the object in the first frame. III.4.2.

Fig. 4. General Flow Diagram for Improved Particle Swarm Optimization

Object Tracking Process

In the below section, the Improved PSO thus presents better solution and the steps engaged are given. The general steps involved in the Improved Particle Swarm Optimization is showed in the Fig. 4.

The tracking process is executed with the assist of the optimization process in the suggested method. The optimization mechanism employed in our suggested Copyright © 2014 Praise Worthy Prize S.r.l. - All rights reserved


643

P. Mukilan, A. Wahi

III.5.1. Steps in Improved Particle Swarm Optimization

viii. The process is replicated till the solution with better fitness value is attained. Accordingly, PSO can be used to work out the optimization problems just like other evolutionary algorithms. It can be employed in dissimilar fields like signal processing, robotics, simulations applications etc. At this moment, in order to select the improved solution for the tracking process we have applied Improved PSO. At last, we collect the output attained from optimization for delivering improved tracked results.

The different steps involved for executing the Improved PSO is described beneath: i. First initialize a population of particles (solutions) with position and velocity selected arbitrarily for ndimension in the problem space. ii. For each of these arbitrarily generated particles assess the optimization fitness functions in nvariables. iii. Then cross over operation is executed in order to produce improved particle. In the cross over process the particles from dissimilar position are exchanged to produce better offspring so that the particle can produce better fitness value compared to the parent particle. iv. Choose the particle with the best fitness value, reinitialize its position. Along with this assess the particle with the worst fitness value, whether its novel position is satisfactory, if it is in an adequate range then revise its position or else a novel position is allocated to the particle arbitrarily in its neighborhood and then mend the position and velocity of other particles by the expression specified beneath:





v j  i  1  v j  i   a1d1 g j  i   h j  i  



 a2 d 2 k j'

i   h j i 

h j  i  1  h j  i   v j  i  1

III.6. Fuzzy C Means Clustering in Order to Cluster the Tracked Result Cluster study is a method for categorizing data, i.e., to separate a specified dataset into a set of classes or clusters. The objective is to split the dataset in such a way that two cases from the same cluster are as similar as feasible and two cases from different clusters are as dissimilar as feasible [22]. To join an attraction of data to clusters with repulsion between different clusters is the plan behind cluster repulsion. Now, the distance between clusters and the data points allocated to them should be minimized. Fuzzy c-means (FCM) is a method of clustering which allows one piece of data to two or more clusters. The degrees of membership to which a specified data point belongs to the dissimilar clusters are calculated from the distances of the data point to the cluster centers regarding the size and the shape of the cluster as uttered by the extra prototype information. The nearer a data point lies to the center of a cluster, the higher is its degree of membership to this cluster. Therefore the problem to separate a dataset    D  d1 ,d 2 ,...,d n  R P into k clusters can be uttered as

(9)

(10)

In the above equations, a1 and a2 signifies the acceleration constants that is required for combining each particle with the pbest ( g j ) and





the task to minimize the distances of the data points to the cluster centers and to maximize the degrees of membership. In probabilistic fuzzy clustering the task is to minimize the objective function:

gbest ( k j ), j is the number of particles (1, 2,…, N), i is the number of iteration, d1 and d 2 are the random values between [0,1], which are employed to keep the diversity of the group particles. v j  i  1 is the velocity of particle j in the  i  1 -

k

f  D, ,  

th iteration, h j  i  is the position of particle j in

n

 

 ijz t 2  i ,d j 

(11)

i 1 j 1

the i th iteration. v. At present compare this fitness value with the particles pbest value. If these present fitness values are better than the pbest then select the present fitness value as the pbest for the additional processing. vi. These fitness values is compared with the overall best earlier values and if the current value is better then revise the gbest for the current particle and value as the new gbest. vii. Modify the velocity and the position of the particle and then do again the steps till the criterion of better fitness is attained. The velocity and the position of the particle are varied with the assist of the Eqs. (1) and 2.

 where ij   0 ,1 is the membership degree of datum d j to cluster ki , i is the prototype of cluster ki , and    t i ,d j is the distance between datum d j and    prototype i .  is the set of all k clusters 1 ,2 ,..., k .





   ij  is called the fuzzy partition matrix and the parameter z is called the fuzzifier. In possibilistic fuzzy clustering, more sensitive assignment of degrees of membership is accomplished by dropping constraint which forces ij away from zero for all i  1, 2,...,k  . That is, the objective function J is adapted to:



644

P. Mukilan, A. Wahi

k

f  D, ,  

 

n

k

n

 ijz t 2  i ,d j     i  1  ij  i 1 j 1

i 1

z

After the segmentation process the feature extraction process is executed where different features from the segmented frames are extorted for tracking process.

(12)

j 1

where,  i  0 . The first term leads to a minimization of the weighted distances while the second term suppresses the trivial solution by making ij  0 for all

IV.

The suggested object detection and tracking system by means of the low level features was put into action in the working platform of MATLAB. The detection and tracking process is checked with dissimilar frames of video and the upcoming result of the suggested work has been demonstrated beneath. At first, the video are segmented to dissimilar shots or frames and next features are extorted followed by the detection and tracking process. Fig. 5(a) is the original image of the car attained from the shot segmentation. Fig. 5(b) is the output after the feature extraction process by motion estimation. Fig. 5(c) demonstrates the segmented output and lastly Fig. 5(d) displays the tracked image of the object in the first frame. Likewise for dissimilar frames the process is repeated and lastly the object is tracked. Fig. 6(a) given beneath is the original image of the car attained from the shot segmentation. Fig. 6(b) is the output after the feature extraction process by motion estimation. Fig. 6(c) demonstrates the segmented output and lastly Fig. 6(d) demonstrates the tracked image of the object in the first frame. Likewise for dissimilar frames the process is replicated and lastly the object is tracked. The suggested methodology proved to be more successful and precise in object detection and tracking.

i  1, 2,...,k  . This strategy is called possibilistic clustering, because the membership degrees for one datum look like the possibility. The formula for revising the membership degrees that is obtained from this objective function is:

ij  t 1   

2



1 1   d j , i  z 1   i 

(13)



 is chosen for each cluster separately and can be determined by the equation: j 

B Ri

 

n

 ij z t 2  d j ,i 

(14)

j 1

where usually, B =1 and Ri 

Results and Discussion

 ij .

(a)

(b)

(c)

(d)

Figs. 5. Results of object tracking in first frame (Video1): (a) Input frame; (b) motion extracted output (c) segmented output; (d) Tracked output



645

P. Mukilan, A. Wahi

(a)

(b)

(c)

(d)

Figs. 6. Results of object tracking in second frame (Video1): (a) Input frame; (b) motion extracted output (c) segmented output;(d) Tracked output

Fig. 7. The final clustered output of tracked object

Fig. 8(a) is the original image of the pot attained from the shot segmentation. Fig. 8(b) is the output after the feature extraction process by motion estimation. Fig. 8(c) demonstrates the segmented output and lastly Fig. 8(d) demonstrates the tracked image of the object in the first frame. Likewise for dissimilar frames the process is replicated and lastly the object is tracked. Fig. 9(a) is the original image of the pot attained from the shot segmentation. Fig. 9(b) is the output after the feature extraction process by motion estimation. Fig. 9(c) demonstrates the segmented output and lastly Fig. 9(d) demonstrates the tracked image of the object in the first frame. Likewise for dissimilar frames the process is replicated and lastly the object is tracked.

V.

Performance Analysis

The accuracy and recall value for the suggested method are worked out for analyzing the performance. Let the object to be tracked be indicated by OT and the tracked output is indicated as TO , then accuracy and recall is stated as: O  TO  precision  T (15) TO 

recall 


OT  TO  OT 

(16)


646

P. Mukilan, A. Wahi

(a)

(b)

(c)

(d)

Figs. 8. Results of object tracking in second frame (Video1): (a) Input frame; (b) motion extracted output (c) segmented output;(d) Tracked output

(a)

(b)

(c)

(d)

Figs. 9. Results of object tracking in second frame (Video1): (a) Input frame; (b) motion extracted output (csegmented output;(d) Tracked output



647

P. Mukilan, A. Wahi

Fig. 10. The final clustered output of tracked object

Precision measures how much of TO covers the OT and recall measures how much of OT is wrapped by the TO . By Eq. (12) and Eq. (13), the precision and recall values for the query image are computed for the suggested method and moreover for the present method. The F-measure for the suggested method is next computed by means of the expression:

 precision  recall  F  2   precision  recall 

(17) Fig. 11. Graphical representation of Average F-measure for proposed and existing method

The values attained from the calculation are specified in Table I. These values are applied for the analysis of performance between the suggested and present method. Here the present method is the earlier paper where object detection and tracking by low level features is executed. Each values connecting to the methods are entered in the table for comparison and from the table it is obvious that our suggested method conveys superior precision and recall than the present method. Now the present method is the vision based object detection and tracking [22].

VI.

Conclusion

We have suggested a competent motion based object detection and tracking system in this document. We improved a distinctive technique where optimization algorithm is applied in order to track the necessary object in the video. As the results demonstrates the suggested methodology proved to be more competent and precise in object detection and tracking than the earlier techniques. To show the efficiency of our suggested technique we have compared the precision and recall value along with F-measure of the suggested technique with present method for the object detection and tracking process. As per the performance study, it is obvious that our suggested technique presents better F-measure value when comparing with other method. Therefore it can be fulfilled that our suggested method is competent in the field of object detection and tracking.

TABLE I PRECISION AND RECALL FOR THE PROPOSED METHOD Performance Analysis F-Measure Precision Recall S.No Proposed Existing Proposed Existing Proposed Existing Method Method Method Method Method Method 1 0.9997 0.71 0.9611 0.15 0.9800 0.2477 2 0.9611 0.65 0.9621 0.24 0.9616 0.3506 3 0.9987 0.59 0.7423 0.35 0.8516 0.4394 4 0.7409 0.52 0.7489 0.46 0.7449 0.4882

The average F-measure value for the proposed and existing method is found out and the corresponding graph is shown in Fig. 11.

References [1]

TABLE II AVERAGE F-MEASURE FOR PROPOSED AND EXISTING METHODS Methods F-measure Proposed Existing Average F-measure 0.8845 0.3815

[2]


Barron, J., Fleet, D., and Beauchemin, S. 1994. Performance of optical flow techniques. International Journal of Computer Vision Vol. 12, No. 4, pp: 43–77, 1994. Xing, Junliang, Haizhou Ai, and Shihong Lao. "Multi-object tracking through occlusions by local tracklets filtering and global tracklets association with detection responses." In Computer Vision and Pattern Recognition. pp: 1200-1207, 2009.


648

P. Mukilan, A. Wahi

[3]

[4]

[5]

[6]

[7]

[8]

[9]

[10]

[11]

[12]

[13]

[14]

[15]

[16]

[17]

[18]

[19]

[20]

[21]

[22]

Okuma, K., Taleghani, A., De Freitas, N., Little, J. J., & Lowe, D. G. A boosted particle filter: Multitarget detection and tracking. In Computer Vision-ECCV, pp: 28-39, 2004 Friedman, J., Hastie, T., and Tibshirani, R. Additive logistic regression: A statistical view of boosting annals of statistics. Vol. 2, pp: 337–374, 2000. Rasmussen, C. and Hager, G. Probabilistic data association methods for tracking complex visual objects. Vol. 23, No. 6, pp: 560–576, 2001. Wren, C., Azarbayejani, A., And Pentland, A. Real-time tracking of the human body. IEEE Trans. Patt. Analy. Mach. Intell., pp: 780–785, 1997. Kim, C., & Hwang, J. N. Fast and automatic video object segmentation and tracking for content-based applications. Circuits and Systems for Video Technology, Vol. 12, No. 2, pp: 122-129, 2002. Mittal, A. and Davis, L. M2 tracker: A multiview approach to segmenting and tracking people in a cluttered scene. Vol 3, pp: 189–203, 2003. Jepson, A., Fleet, D., And Elmaraghi, T. Robust online appearance models for visual tracking. IEEE Trans. Patt. Analy. Mach. Intell. Vol. 25, No. 10, pp: 1296–1311, 2003. Paragios, N., & Deriche, R. Geodesic active contours and level sets for the detection and tracking of moving objects. Pattern Analysis and Machine Intelligence, Vol. 3, pp: 266-280, 2000. Vu, T. D., Burlet, J., & Aycard. Grid-based localization and local mapping with moving object detection and tracking. Information Fusion, Vol. 12, No. 1, pp: 58-69, 2007. Hsia, C. H., & Guo, J. Ming. Efficient modified directional lifting-based discrete wavelet transform for moving object detection. Signal Processing, pp: 138-152. Cuevas, C., & García. Improved background modeling for realtime spatio-temporal non-parametric moving object detection strategies. Image and Vision Computing. 2013. Karasulu, Bahadir, and Serdar Korukoglu. "Moving object detection and tracking by using annealed background subtraction method in videos: Performance optimization." Expert Systems with Applications, Vol. 1, pp: 33-43, 2012. Chieh-Chih Wang and Chuck Thorpe, "Simultaneous Localization and Mapping with Detection and Tracking of Moving Objects" In.Proc.of the IEEE International Conference on Robotics and Automation (ICRA), 2002. S. Kirkpatrick,C. D. Gelatt and M. P. Vecchi, "Optimization by Simulated Annealing Science, New Series, Vol. 220, No. 4598,pp. 671-680,1985 Zhaozheng Yin and Robert T. Collins, ”Object Tracking and Detection after Occlusion via Numerical Hybrid Local and Global Mode-seeking” , IEEE Computer Vision and Pattern Recognition (CVPR'08), Anchorage, Alaska, 2008. Leibe, B.; Schindler, K.; Van Gool, L., "Coupled Detection and Trajectory Estimation for Multi-Object TrackingComputer Vision, 2007. ICCV 2007. IEEE 11th International Conference on, 14-21 Oct. 2007, pp.1-8. J. Fernandeza, R. Guerreroa, N. Mirandaa and F. Piccolia ,"MultiLevel Paralelism In Image Identification,"Mecanica Computational,Vol.28,pp.227-240,Argentina,Nov 2009. Kalpesh R Jadav, Prof.M.A.Lokhandwala and Prof.A.P.Gharge ,"Vision based moving object detection and tracking", National Conference on Recent Trends in Engineering & Technology, May 2011. Qiu, Y., Liu, C., A collaborative task allocation mechanism for wireless sensor networks, (2012) International Review on Computers and Software (IRECOS), 7 (7), pp. 3821-3825. Ma, Z., Wang, Y., Zheng, Y., Zou, X., An improved segmentation method based on Semi-Fuzzy cluster, (2012) International Review on Computers and Software (IRECOS), 7 (7), pp. 3452-345.

Authors’ information P. Mukilan obtained his Bachelor’s degree in Electrical and Electronics Engineering from University of Madras. Then he obtained his Master’s degree in Applied Electronics from Anna University and Pursuing PhD in Faculty of Electrical majoring in Video Processing from Anna University.Currently, he is a Associate Professor and Head at the Faculty of Electronics and Communication Engineering, C.M.S.College of Engineering and Technology. His Specializations include Image Processing, Pattern Recognition and Computer Graphics. His current research interests are Segmentation, Video Processing , Object Detection and Recognition. Dr. A. Wahi received the B.Sc.degree from the LNM University,Darbhanga,India,in 1988,the M.Sc. degree from the LNM University, Darbhanga, India, in 1991,and the Ph.D degree from the Banaras Hindu University(BHU), Varanasi, India,in 1999.He is currently an Professor in the Deparment of Information Technology at Bannari Amman Institute of Technolgy.His research interests are in Neural Networks, Fuzzy Logic and Pattern Recognition.He is a member of Computer Society of India and a Life member of ISTE.



649


Prediction Algorithms for Mining Biological Databases Lekha A.1, C. V. Srikrishna2, Viji Vinod3 Abstract – The paper attempts to understand the efficiency of prediction algorithms on breast cancer data using two kinds of experiments. The experiments are based on categorical data and quantitative data. It attempts to propose a predictive model that can be used to predict the risks of breast cancer. The paper suggests possible ways of improving the efficiency of the predictive model for a given dataset. The study reveals that the efficiency of a mining algorithm is a function of many variables of the dataset. The study proposes a predictive model through a case study. Copyright © 2014 Praise Worthy Prize S.r.l. - All rights reserved.

Keywords: Biological Databases, Breast Cancer, Efficiency Predictive Algorithms, Statistical Analysis

I.

Introduction

II.

In all the spheres of human life many new fields of study emerged with the inception of Information Technology [1]-[30]. Bioinformatics is a field that has attracted many researchers due of its importance and influence in the modern world [3]. The introduction of computer science in the field of medicine has paved the way for many new aspects in the medical sciences like Genomics, Proteomics, and Cell Biology. These concepts’s existences were difficult to believe in the past [9]. Mining biological databases is an option that transformed the medical science from its traditional reactive methods to modern proactive approach. The biological databases contain information about the life sciences [16] and are enriched with the research based knowledge of proteomics, metabolomics, genomics, phylogenetics [19] and microarray sciences. The information present in the database is the real and true information though it can be current or old. The databases contain both historical data and the current facts [21]. The database can also contain information about genes and protein sequence [24]. Mining methods have enabled people to extract facts from the database in a way such that it provides altogether a new set of information. This information cannot be easily observed by human beings in generalized reports. Data mining is also known as Knowledge Discovery in database due to the fact that it helps in bringing out the hidden facts [24]. The new set of information assists the decision makers and researchers [20] in focusing the unexplored pattern or information of the field. Mining takes place with the help of artificial intelligence i.e. which explores information to extract new patterns. The extracted information presents a perspective which is not considered so far.

Prediction through Mining

Prediction of the likely biological patterns in the species needs mining of biological databases. It helps the medical science identify the risk of potential diseases in the generations to come. Researchers use mining techniques to prepare the treatment and medication for the likely disease. Prediction is a proactive approach that can save further generations from painful sufferings of diseases for whom treatment is made possible, only through the prediction practices. In case the predictions are not made, the medical practitioners will not be able to diagnose the problem in patient or the prescription may be wrong and the condition of patient is bound to worsen. Mining of data analyses the information in unique ways that helps in predicting future trends and likely patterns that can occur [9]. The point of concern is the validity of the predicted information that is largely determined by the efficiency of prediction algorithm. In bioinformatics, prediction is made through a well defined series of steps [10]. The algorithm contains the set of rules and procedures that are followed in the information analysis process. It is mandatory that the underlying algorithm takes into consideration all the possible variables and their interactions in order to predict reliable information, [11] which can cause changes in the predicted behaviour [12].

III. Statistical Analysis of Mining Algorithm Prediction is usually based on historical data. Statistical methods are used in analyzing the historical data. These methods are normally embedded in the algorithms [13] and increase the efficiency of prediction algorithm [14]. The statistical measure of correlation throws light upon the type of association between variables and the strength of that relationship but it does



650

Lekha A., C. V. Srikrishna, Viji Vinod

not reveal the causality [15]. It is important to measure the causality in mining biological databases, where major health related developments are to be made based on the predictions. By using the linear regression model, the prediction is made easy but the accuracy of predicted value is highly questionable due to approximations and assumptions [17], [18]. The prediction on the basis of one variable tends to a misleading prediction. In reality, there are always multiple independent variables leading to change in the values of dependent variables. It is crucial to analyse the combined effect of the variables involved so that the most influential one can be controlled which warrants for a better statistical model like multiple correlation and, or multiple regression analysis to predict the current and future trends. The method usually employed to predict the value of an unknown variable is the regression analysis. The prediction is made on historical data only and it may be possible that the value predicted may be misleading. If a prediction model is applied to generate the future trends [19] and the validity of the predicted values increase [7] then the algorithm is considered as efficient. The prediction model includes the factors like environment, possible changes in the past data and the context in which the predicted values will be used. For better prediction, the context in which past data is gathered also plays a vital role [20]. The encompassing intelligence analyses the data accordingly and presents the trend which is workable for future decisions.

IV.

We consider two cases to make predictions – (i) classification by considering data as categorical (ii) classification by considering data as numeric and quantification is done by considering an identifier [15].

VI.

The main aim of the first experiment is to predict the class by finding the efficiency of different algorithms considering the data as categorical. The experiment predicts whether the data being tested can be correctly predicted and classified as benign or malignant class. The experiment was completed with five different algorithms namely Decision Tree, OneR, PART, JRip and ZeroR. Four different predictive methods – Cross Validation, Percentage Split, Testing data and Training data are used in this analysis. For the cross validation the method used is 10 fold. While using percentage split the number of instances gets reduced to 238 for the particular case study of breast cancer since it uses the 2/3rd method. The mean absolute error for each of the algorithm was measured as it is one of the widely used statistics for regression. The following table details the mean absolute error for all the algorithms. TABLE I MEAN ABSOLUTE ERROR FOR CASE STUDY

Cross validation Percentage Split Training set Testing Set Difference

Efficiency of the Algorithm

The prediction model contains the built in mechanism to create a sequence of value observed over the period of time. The values are analysed with respect to the various identified variables. The algorithm with inbuilt intelligence uses multiple dimensions of value analysis. The algorithm creates multiple sequences based on complex computational methodology and presents the report to the decision maker. The complexity of the algorithm also determines the result’s reliability [21]. The analysis of sequential dataset is the most common feature of prediction algorithms applied in the present era [22]. To support the claims made earlier we consider case studies in breast cancer.

V.

Experiment 1

ZeroR

OneR

JRip

PART

0.452

0.073

0.0618

0.0685

Decision Table 0.0865

0.458

0.0798

0.06

0.0627

0.0963

0.4519 0.4519 0.0061

0.073 0.0787 0.0068

0.039 0.0433 0.0228

0.0277 0.0703 0.0302

0.0661 0.0339 0.0458

From Table I it can be seen that for all the algorithms the error is quantifiable. It is obvious that for a best accuracy we wish to obtain the smallest possible value for the error. ZeroR has a consistent higher value of error for all the methods compared to the others. It is observed that variations between the methods are not significant for any given algorithm. The following method is being proposed to get a clear picture of the efficiency of the above said algorithms. A confusion matrix (Table II) is considered for each of the above algorithms that shows (i) True Positive (TP), (ii) True Negative (TN), (iii) False Negative (FN) and (iv) False Positive (FP). In medical field a false positive is the cause of unnecessary worry or treatment whereas a false negative gives the patient the dangerous illusion of good health and the patient might not go for an available treatment.

Input

The analysis of the efficiency of predictive mining algorithm on the data set related to breast cancer. The breast cancer domain was obtained from the University of Wisconsin Hospital, Madison, Wisconsin, USA. It was created by Dr. William H. Wolberg (physician) and donated by Olvi Mangasarian [25]. This data has 699 instances described by 9 attributes + one class attribute. The set includes 458 instances of one class and 241 instances of another class.

TABLE II CONFUSION MATRIX Actual value Predicted TP FP Outcome FN TN Total P N



651


TABLE IV CROSS VALIDATION ANALYSIS

In the following paragraphs the meaning of each term is given for clarity and ability. True Positive denotes the number of benign events of cancer correctly classified as benign events in the study. False positive denotes the number of malignant events of cancer incorrectly classified as benign events in the study. True negative denotes the number of malignant events of cancer correctly classified as malignant events in the study. False negative denotes the number of benign events of cancer incorrectly classified as malignant events in the study. Different measures are calculated using these values. The measures on which the prediction algorithm is analyzed are (i) number of correctly classified and incorrectly classified instances, (ii) Accuracy, (iii) Sensitivity, (iv) Specificity, (v) Positive Predictive Value (PPV), (vi) Negative Predictive Value (NPV). The meaning of each parameter is given below for immediate reference. Sensitivity measures the proportion of the actual positives that are correctly identified as true. Specificity measures the proportion of the negatives that are correctly identified as false. A perfect predictor should have 100% sensitivity and 100% specificity. A high PPV means that the predictor always classifies correctly but has to be taken in account with the value of NPV since predictive values are inherently dependent upon the prevalence of both true and false. NPV is a numerical value for the proportion of individuals with a negative test result who are free of the target condition—i.e., the probability that a person who is a test negative is a true negative. A high NPV means that the predictor rarely misclassifies. We consider the four different methods to assess the parameters listed above and present the calculated results in the form of tables for the case study in the following paragraphs.

ZeroR Time taken to build model (in secs)

14

22

19

17

241

37

10

24

16

0

204

231

217

225

0.08

0.03

Decision Table 0.16

656

666

43

33

0.94 0.95 0.92 0.96

0.95 0.96 0.93 0.96

0.90

0.93

0.8632

0.8956

PART

From Table IV it can be noted that the time taken by Decision Table to build the model is the highest and the time taken by ZeroR is the least. The ratio between the number of correctly classified and incorrectly classified instances is very high. ZeroR has the highest accuracy with 100%. All the other algorithms have a very high value of accuracy of approximately 95%. The Kappa statistic result shows that there is no complete agreement with the true class. All the algorithms have nearly 96% sensitivity which is quite high. The test is able to detect 96% of the people with the correct cancer and misses 4% of the people. The algorithms differ in the values of the specificity with OneR having the highest specificity of 94%. It means 14 persons out of 51 persons with negative results are truly negative and 37 individuals test positive for a type of cancer disease which they do not have. ZeroR has the highest PPV that means that it correctly classifies the class as recurrent class. JRip has the highest NPV of 96%. ZeroR has a 0 NPV which means that the algorithm has the highest probability of misclassification. The value of kappa statistic is high ranging from 0 to 0.8999 for all algorithms which indicates chance agreement.

TABLE III CROSS VALIDATION CONFUSION MATRIX DETAILS ZeroR OneR JRip PART Decision Table 458 444 436 339 441 0

0.01

JRip

Correctly 458 648 667 classified instances Incorrectly 241 51 32 classified instances Accuracy 1.00 0.93 0.95 Sensitivity 1.00 0.92 0.98 Specificity * 0.94 0.91 Positive 1.00 0.97 0.95 Predictive Value Negative 0.00 0.85 0.96 Predictive Value Kappa Statistic 0.00 0.8348 0.8999 * - cannot be calculated as denominator is zero

Using Cross validations

True Positive True Negative False Positive False Negative

0

OneR

Percentage Split – 66% TABLE V PERCENTAGE SPLIT CONFUSION MATRIX DETAILS Decision ZeroR OneR JRip PART Table True Positive 152 146 146 145 146 True Negative 0 6 6 7 6 False Positive 86 13 4 3 6 False 0 73 82 82 80 Negative

From Table III it is observed that PART has the lowest number of correctly classified instances. The FP value of ZeroR is the highest for all the algorithms. It means that there are 241 cases when a malignant event is incorrectly classified as benign event. The ZeroR algorithm has a 0 value for FN. The JRip has the highest value of 231 for FP. This leads to a dangerous situation where 231 benign events are incorrectly classified as malignant events. Using Table III we get the results of the following table.

From Table V it can be seen that all the algorithms except ZeroR have a near consistent number of correctly classified instances.



652


TABLE VIII TRAINING SET ANALYSIS

The FP value of ZeroR is 86 and is the highest for all the algorithms. This causes a dangerous misclassification since a malignant event is incorrectly classified as benign event. The ZeroR algorithm has a 0 value for FN. Using Table V we get the results of the following table.

ZeroR

OneR

JRip

Time taken to build 0 0.01 0.1 model (in s) Correctly classified 152 219 228 instances Incorrectly classified 86 19 10 instances Accuracy 0.64 0.92 0.96 Sensitivity 0.64 0.92 0.97 Specificity * 0.92 0.93 Positive Predictive 1.00 0.96 0.96 Value Negative Predictive 0.00 0.85 0.95 Value Kappa Statistic 0.00 0.8239 0.9094 * - cannot be calculated as denominator is zero

0.04

Decision Table 0.14

227

226

11

12

0.96 0.98 0.92 0.95

0.95 0.96 0.93 0.96

0.96

0.93

0.9006

0.8908

PART

0.03

Decision Table 0.21

688

674

11

25

0.98 0.99 0.97 0.98

0.96 0.97 0.96 0.98

0.98

0.94

0.9653

0.9204

PART

From Table VIII it can be observed that the time taken by Decision Table to build a model is the highest. The time taken by ZeroR and OneR is the least. The ratio between the number of correctly classified and incorrectly classified instances is very high. JRip and PART have the highest accuracy of 98%. All the algorithms have more than 92% sensitivity except ZeroR that has 66%. JRip has the highest at 100%. It means that the actual positives correctly identified are all true. All algorithms have above 90% specificity except ZeroR. PART has the highest specificity with 97%. Though ZeroR has 100% PPV it has 0% NPV which means that the algorithm has the highest probability of misclassification. JRip has the highest NPV. The Kappa statistic result shows that there is no complete agreement with the true class but nearing to perfect agreement.

From Table VI we can ascertain that the time taken by Decision Table to build a model is the highest. The time taken by ZeroR and OneR is the least. The ratio between the number of correctly classified and incorrectly classified instances is very high. Since the method uses only 1/3rd of the instances the total number of instances is reduced to 233. JRip and PART have the highest accuracy with 96%. All the algorithms except ZeroR have approximately 95% sensitivity. All the algorithms have nearly 92% specificity except ZeroR. Though ZeroR has 100% PPV it has 0% NPV which means that the algorithm has the highest probability of misclassification. PART has the highest NPV. The value of kappa statistic is very high ranging from 0 to 0.9094 for all algorithms which indicates near complete agreement.

Using Testing data TABLE IX TESTING SET CONFUSION MATRIX DETAILS

Using Training set TABLE VII TRAINING SET CONFUSION MATRIX DETAILS


JRip

Time taken to 0 0 0.08 build model (in secs) Correctly 458 648 684 classified instances Incorrectly 241 51 15 classified instances Accuracy 0.66 0.93 0.98 Sensitivity 0.66 0.92 1.00 Specificity * 0.94 0.95 Positive 1.00 0.97 0.97 Predictive Value Negative 0.00 0.85 0.99 Predictive Value Kappa Statistic 0.00 0.8348 0.953 * - cannot be calculated as denominator is zero

TABLE VI PERCENTAGE SPLIT ANALYSIS ZeroR

OneR

ZeroR

OneR

JRip

PART

458 0

444 14

445 13

451 7

Decision Table 448 10

241

37

2

4

15

0

204

239

237

226


ZeroR

OneR

JRip

PART

458 0

442 16

445 13

451 7

Decision Table 448 10

241

39

5

9

17

0

202

236

232

224

The Table IX illustrates that ZeroR has the lowest number of correctly classified instances. The FP value of ZeroR is 241. This causes a dangerous misclassification since a malignant event is incorrectly classified as benign event. The ZeroR algorithm has a 0 value for FN. All the algorithms have a considerable higher value of FN. Using Table IX results found are in Table X. The testing data method uses all the instances in the testing set. From Table X it is noted that the time taken by Decision Table to build a model is the highest.

From Table VII it can be noted that ZeroR has the lowest number of correctly classified instances. The FP value of ZeroR is 241 and is the highest for all the algorithms. This causes a dangerous misclassification since a malignant event is incorrectly classified as benign event. The ZeroR algorithm has a 0 value for FN. Using Table VII the following results are got.



653


TABLE X TESTING SET ANALYSIS ZeroR

OneR

JRip

Time taken to 0 0.03 0.17 build model (in s) Correctly 458 644 681 classified instances Incorrectly 241 55 18 classified instances Accuracy 0.68 0.92 0.97 Sensitivity 0.66 0.92 0.99 Specificity * 0.93 0.95 Positive 1.00 0.97 0.97 Predictive Value Negative 0.00 0.84 0.98 Predictive Value Kappa Statistic 0.00 0.8218 0.9434 * - cannot be calculated as denominator is zero

0.15

Decision Table 0.21

683

672

16

27

0.98 0.98 0.97 0.98

0.96 0.96 0.96 0.98

0.96

0.93

0.9492

0.9139

PART

embedded in the algorithm and the statistical methods used in prediction process play an important role in the prediction of data.

VII.

Experiment II

The main aim of the second experiment is to predict by quantifying an identifier. Nine attributes are considered in this data set to predict whether it is benign or malignant cancer. These attributes assume values from [1]-[10] [24]. The attributes are divided into two intervals [1]-[5] and [6]-[10]. If the entire range is considered it is equivalent to applying the different algorithms already experimented upon in the previous analysis. For immediate and easier reference each of the attributes are assigned numbers and given below. TABLE XI ATTRIBUTE NUMBER ASSOCIATION Attributes Number Clump_thickness 1 Cell_Size_uniformity 2 Cell_Shape_uniformity 3 Marginal_Adhesion 4 Single_Epi_Cell_Size 5 Bare_Nuclei 6 Bland_Chromatin 7 Normal_Nucleoli 8 Mitoses 9

The time taken by ZeroR is the least. The ratio between the number of correctly classified and incorrectly classified instances is very high. PART has the highest accuracy with it being 97%. All algorithms have approximately 92% and above sensitivity except ZeroR with 66% sensitivity. JRip has the highest with 99%. All algorithms have above 93% specificity except ZeroR. PART has the highest specificity. Decision Table and PART have the highest PPV with 98%. Though ZeroR has 100% PPV it has a 0 NPV which means that the algorithm has the highest probability of misclassification. JRip has the highest NPV with 98%. The Kappa statistic result shows that though there is no complete agreement with the true classes it nears the complete agreement with PART algorithm having 0.9492. From the above tables it is found that for ZeroR algorithm specificity cannot be calculated. This means that the algorithm has the highest probability of misclassification. The algorithm also has the lowest number of FN value for all the four methods. It is also found that the Kappa Statistic is the highest for the method of training data set. It is has a nearly 96% agreement with the true class. ZeroR has a 0% agreement with the true class. OneR, JRip algorithms have around 80% agreement with the true class. The number of correctly classified instances is considerably more than the number of incorrectly instances using all the methods for all algorithms except ZeroR. JRip, PART and Decision Table algorithms have a consistent high sensitivity value using all the four models. The analysis of the experimental results leads to the understanding that the steps of algorithm, the models

The individual attribute values for the given range distribution is as follows. From Table XII it is understood that for all the attributes the maximum number of instances is assumed in the interval [1]-[5]. This data is further analyzed to understand when they form benign cancer or malignant cancer. It is consolidated and placed in the Table XIII. From such table it is noted that when the values of the attributes is [1]-[5] most of the instances form benign cancer whereas when the interval is [6]-[10] most of the instances form malignant cancer. The number of benign cancer instances when the interval is [1]-[5] is nearly the same. The number of malignant cancer instances when the value is in [6]-[10] varies from 32 to 164. The relevance of dividing the attributes values is observed from the table. From Table XIII it is seen that the number of benign cancer instances are nearly the same for all the attributes. Contribution from each attribute is the same and it is not possible to identify a particular attribute in the case of benign cancer.

TABLE XII ATTRIBUTE RANGE ASSOCIATION Attributes Range 1-5 6-10

1

2

3

4

5

6

7

8

9

513 186

550 149

547 152

579 120

592 107

509 190

557 142

559 140

665 34



654


TABLE XIII ATTRIBUTE-RANGE-CLASS ASSOCIATION Attributes Class Benign Malignant Benign Malignant

1

2

3

4

5

6

7

8

9

Interval

436 77 22 164

452 98 6 143

451 96 7 145

451 128 7 113

449 143 9 98

438 71 6 168

449 108 9 132

446 113 12 128

456 209 2 32

1-5

Table XIV gives the details of the severity of the attributes towards benign or malignant cancer. It can be seen from the table that the ratio of benign to malignant is in the range [1]-[5] ranges from 6:1 to 2:1. The attribute #9 – Mitoses plays an important role in identifying malignant cancer when its values are in the interval [6]-[10]. Research has shown that this is the attribute that plays an important role in turning a normal cell to cancerous cell [25], [26]. From Table XV it is seen that attribute #6 – Bare_Nuclei and attribute #1 – Clump_thickness play an important role in identifying malignant cancer when its range of value is in the interval [6]-[10]. Only Attribute #1 plays an important role in identifying benign cancer when its value is in the interval [1]-[5]. Research shows that Nucleus and Clump_thickness plays an important role in converting a cell into a malignant cell [27], [28].

IX.

6-10

Results of Experiment I

Graph 1 shows that ZeroR has a consistently high mean absolute error of 0.45. The other algorithms have a better mean absolute error (MAE). For prediction it is always preferrable to have a low value for the mean absolute error. It is difficult to determine a better predictive algorithm based on only the MAE statistics. We consider the other measures to determine the best predictive algorithm. The analysis of Graph 2 illustrates that for all the algorithms except ZeroR, the accuracy is nearly 90% for all methods. For ZeroR, the accuracy varies from 60% to 100%. The analysis of Graph 3 depicts that OneR and Decision Table have a consistent sensitivity rate of 92% and 96% respectively for all the methods. For the other two algorithms it ranges from 95% to 100%. ZeroR has an inconsistent sensitivity ranging from 60% to 100%.

VIII. Results and Discussions This paper deals with finding efficiency of an algorithm given a large dataset. Five algorithms have been considered namely Decision Tree, PART, OneR, JRip and ZeroR and four different predictive methods of cross validation, percentage split, inducing error. The mean absolute error for each of the algorithms in each method is considered first. Further we have considered the confusion matrix for the calculation. The measures on which the prediction algorithms are analyzed are (i) number of correctly classified and incorrectly classified instances, (ii) accuracy, (iii) sensitivity, (iv) specificity, (v) positive predictive value (PPV), (vi) negative predictive value (NPV). In what follows we represent the calculated results in the form of graphs.

Graph 1 – MAE Statistics

TABLE XIV ATTRIBUTE-CLASS ASSOCIATION BASED ON SEVERITY Attributes 9 2 3 4 5 7 8 6 Class Benign 456 452 451 451 449 449 446 438 Malignant 209 98 96 128 143 108 113 71 TABLE XV ATTRIBUTE-CLASS ASSOCIATION BASED ON PRIORITY 2 Attributes 6 1 3 2 7 8 4 5 Class Benign 6 22 7 6 9 12 7 9 Malignant 168 164 145 143 132 128 113 98


1 436 77

9 2 32


655


than 90%. ZeroR has the best PPV value of 100%. OneR is consistent with three methods of Cross Validation, Training Set and Testing Set having a value of 97%. The other three algorithms are consistent with two methods.

Graph 2. Accuracy Statistics

Graph 6. NPV Statistics

The analysis of Graph 6 depicts that Decision Table has the most consistent NPV value of all the algorithms with respect to all the methods. All the methods except ZeroR have an NPV value of 80% and above. ZeroR has a NPV value of 0% which shows that it has the highest probability of misclassification.

Graph 3. Sensistiviy Statistics

Graph 4. Specificity Statistics Graph 7. Kappa Statistics

Graph 4 illustrates that Decision Table has a consistent specificity rate for three methods except Cross Validation. All the other algorithms – OneR, JRip and PART have consistent specifity value for any two methods. The specifity value of all the algorithms except ZeroR in all the methods is above 90%. ZeroR’s specificity measure cannot be calculated.

From the analysis of Graph 7 it is observed that all the algorithms except ZeroR have a Kappa value above 80% for all the methods. ZeroR has a Kappa value of 0 which shows that there is no complete agreement to the true class of prediction. OneR has a consistent kappa value of around 80%. All the other three algorithms have a nearly 90% agreement with the true class.

X.

Results of Experiment II

The analysis of the second experiment based on extracting indicator attributes results in Mitoses and Bare_Nuclei being identified as the indicatore attributes. The literature availableon this tells that mitotic index could stratify women into groups with high and low risk of recurrence. As per National Institutes of Health report gene's Position in the Nucleus can be used to distinguish Cancerous from Normal Breast Tissue.

Graph 5. PPV Statistics

From the analysis of Graph 5 it is noted that all the algorithms have a consistently good PPV value of more



656


XI.

Conclusion

References [1]

From the analysis of the results it is found that PART algorithm categorises the data to the true class better than the other algorithms in case study 1 where the data is categorical and quantitative. The method of cross validation yields the most consistent accurate result for all the different algorithms used for prediction. ZeroR algorithm results in 0% specificity and 0 NPV which means that this algorithm has the highest probability of misclassifications. In cross validation method Decision Tree algorithm has the best accuracy and specificity. PART algorithm has the best accuracy in percentage split method and in testing set method but not the best specificity. In training set method PART algorithm has the highest accuracy and specificity. When the data set consists of numerical values it is found that JRip, PART and Decision Table categorises the data nearly to the true class. All the methods yield the most consistent accurate result for all the algorithms. All the algorithms except ZeroR have a consistent sensitivity rate for all the methods. Decision Table has a consistent specificity rate of 95% for all methods except cross validation. ZeroR algorithm results in 0% specificity and 0 NPV which means that this algorithm has the highest probability of misclassifications. All the algorithms have a nearly 94% and above PPV value. All though the results of both the case studies are presented by considering the parameters (i) number of correctly classified and incorrectly classified instances, (ii) accuracy, (iii) sensitivity, (iv) specificity, (v) positive predictive value (PPV), (vi) Negative predictive value (NPV), (vii) Kappa Statistic the numerical computations based on error reveals that for all the algorithms except ZeroR the mean absolute error is not uniform. The mean absolute error of ZeroR algorithm has a consistent mean absolute error which is around 0.45 which shows that the algorithm is not a good predictor. The second experiment enables in deducing a classification rule for a data set that consists of both categorical and numerical data. When the data set is completely numeric in nature identifier attributes are identified. From the above analysis it is understood that the efficiency of a mining algorithm is found to be the function of many variables such as dataset consisting of huge historical information, the perspective of collected information, context in which predicted information will be used. The steps of algorithm, the models embedded in the algorithm and the statistical methods used in prediction process also play an important role. Prediction model is not a linear model to the present case study. It is also noted that reliable results can be produced if the mentioned points are carefully analyzed in the algorithm design.

[2] [3] [4] [5] [6] [7] [8]

[9]

[10]

[11] [12] [13] [14] [15]

[16] [17]

[18]

[19] [20] [21]

[22] [23]

[24] [25] [26] [27]

Acknowledgments The authors thank the principal and management of PESIT and Dr. MGR Educational and Research Institute for their continued support.

[28] [29]


Anagnostopoulos. Path Prediction through Data Mining. ICPS, n.d. Web. 29 March 2012. Anna Batistatou, “Mitoses and Cancer”, Medical Hypotheses, Volume 63, Issue 2, 2004, Pages 281-282, ISSN 0306-9877, Avery, John. Information Theory and Evolution. USA: World Scientific Publishing Co. 2003. Bartz-Beielstein, Thomas, et al. Experimental Methods in Algorithm Design and Analysis. USA: Springer. 2010. Bertsimas, Dimitris, et al. “Algorithm Prediction of Health-Care Costs.” Operations Research 56. 6 (2008): 1382-1392. Bonate, Peter. Pharmacokinetic-Pharmacodynamic Modeling and Simulation, USA: Springer. 2011. Crawley, Michael. Statistics: An Introduction Using R. USA (John Wiley & Sons. 2011). David Weatherall, Brian Greenwood, Heng Leng Chee and Prawase Wasi, Science and Technology for Disease Control: Past, Present, and Future, Disease Control Priorities in Developing Countries Djahantighi, Farhad et al. “An Effective Algorithm for Mining User behavior in Time-Periods” European Journal of Scientific Research 40. 10 (2010): 81-90. Golriz Amooee, Behrouz Minaei-Bidgoli, Malihe BagheriDehnavi. “A Comparison Between Data Mining Prediction Algorithms for Fault Detection (Case study: Ahanpishegan co.)” 2012. Web. 29 April 2012. Hof, Paul. System Identification 2003. UK: Elesevier Ltd. 2004. http://www.cs.iastate.edu/~cs472/labs/breast-cancerwisconsin.arff http://www.hakank.org/weka/BC.arff http://www.nih.gov/news/health/dec2009/nci-07.htm Lekha A, Srikrishna C V, Vinod Viji, Efficiency of Prediction Algorithms for Mining Biological Databases, IOSR Journal of Computer Engineering (IOSRJCE) ISSN: 2278-0661, ISBN: 2278-8727 Volume 6, Issue 1 (Sep-Oct. 2012), PP 12-21 Leser, Ulf, et al. Data integration in the Life Science. USA: Springer. 2006. Michalski,R.S., Mozetic,I., Hong,J., & Lavrac,N. (1986). “The Multi-Purpose Incremental Learning System AQ15 and its Testing Application to Three Medical Domains.” In Proceedings of the Fifth National Conference on Artificial Intelligence, 10411045, Philadelphia, PA: Morgan Kaufmann. Narasimha Aparna, Vasavi B, Harendra Kumar M L, ” Significance of nuclear morphometry in benign and malignant breast aspirates”, IJABMR, Year : 2013 , Volume: 3 , Issue Number: 1 , Page: 22-26 Pandey, Hari. Design Analysis and Algorithm. New Delhi: University Science Press. 2008. Paradis, Emmanuel. Analysis of Phylogenetics and Evolution with R.USA: Springer. 2011. Ramaswamy, Sridhar, et al. Efficient Algorithms for Mining Outliers from Large Data Sets. The Pennsylvania State University, 2010. Web. 29 April 2012. Rob, Peter, et al. Database systems: design, implementation and management. USA (Cengage Learning, 2009). Rossiter. An Introduction to Statistical Analysis. Netherlands: International Institute of Geo Science and Earth Observation. 2006. SDART. Software Design and Research Technology Ltd., n.d. Web. 29 April 2012. Selzer, Paul. Applied Bioinformatics: An Introduction. USA: (Springer. 2008). Sen, Zekai. Spatial Modeling Principles in Earth Sciences. USA: (Springer. 2009). S J Done, NA Miller, Shi W Wei, M Pintilie, DR McCready, F-F Liu, and A Fyles,” Mitotic Component of Grade Can Distinguish Breast Cancer Patients at Greatest Risk of Local Relapse”, Cancer Research: December 15, 2012; Volume 72, Issue 24, Supplement 3, doi: 10.1158/0008-5472.SABCS12-P2-10-33 Technet. Microsoft, n.d. Web. 29 April 2012. Yan, Xin, et al. Linear Regression Analysis: Theory and Computing. USA: (World Scientific Publishing Co. 2009).


657


[30] Tabatabaee-Y., H., Mehrnejad, M., Kazem Shekofteh, S., Cancer detection based on experimental sampling by genetic-fuzzy classification system, (2012) International Review on Computers and Software (IRECOS), 7 (3), pp. 1062-1069.

Authors’ information 1

Research Scholar, Dr M G R Educational Research Institute, Chennai, India – 600095, Assistant Professor, Department of MCA, PESIT, Bangalore. 2

Professor, Department of MCA, PESIT, Bangalore – 560085.

3

Head, Department of MCA, Dr. MGR Educational and Research Institute, Chennai, India-600095. A. Lekha was born in Bangalore, India in 1979. She received her Bachelor of Science degree from Bangalore University, Bangalore, India and Masters Degree in Computer Applications from Visvesvaraya Technological University, Belgaum in 1998 and 2001 respectively. After working as a lecturer in BMS Institute of Technology, Bangalore, she is currently working as assistant professor in the MCA department of P E S Institute of Technology, Bangalore. She is currently pursuing the Ph.D degree in data mining from Dr. M G R Educational and Research Institute. Her research interests include clustering, classification of data for bioinformatics, programming. C. V. Srikrishna received his PhD degree from Bangalore University in the year 2001. He has been in the field of teaching from the past 25 years. He has published more than 50 papers in reputed national and international journals. He has organized various national conferences and seminars on cryptography, parallel computing, relational algebra, calculus and its applications in DBMS, Discrete Structures and Computer Applications, Networks and Security. He has been working on projects funded by DST, UGC and AICTE. His research interests include modelling and simulation, algorithms to heat transfer and applied mathematics and statistical applications to data mining and industrial process. Viji Vinod started her career in working on real time software engineering. She joined as a software engineer at Pentafour Communications Ltd, Chennai, India in 1999 where she achieved proficiency in developing enterprise resource planning applications in client /server architectures. She has worked as a software developer with Lawrence and Associates ltd, a US based software company where she developed various web based applications. She joined as the faculty of the Department of Information Technology at Dr. M G R University in 2002, becoming senior lecturer of information technology management and computer science. She was the first professor in object oriented analysis and design and component based technology at Dr. M G R University in 2004. She served as an Asst. Professor of Department of Computer Applications at Dr. M G R University from 2006 to 2009, and as Professor and Head of the department of the department of Computer Applications at Dr. M G R Educational and Research Institute, University. She specializes in areas of information technology management and various software engineering issues with the goal of improving both software engineering practices and processes. She has edited one book and published articles and notes in professional journals and conferences.



658


A Grid-Based Algorithm for Mining Spatio-Temporal Sequential Patterns Gurram Sunitha, A. Rama Mohan Reddy Abstract – Relentless accumulation of spatio-temporal data at micro-granularities of time got conceivable with the advent of real-time applications and technologies. Many ubiquitous services such as environmental monitoring, impact assessment, real-time surveillance and navigation support have advanced alongside the advancing technology. All of them require the handling of immensely colossal volumes of spatio-temporal data and withal, the fortification of useful knowledge for real-time decision-making. Mining sequential patterns from spatio-temporal databases is one of the prominent strategies for discovering causal relationships in spatiotemporal data. Sequential patterns give more preponderant insight into the spatial and temporal aspects of different event types and the interactions between them. In this paper, an efficient algorithm has been proposed to mine sequential patterns. Ultimately, the experimentation and comparison of our proposed algorithm with STS-miner algorithm, proved the efficiency of our algorithm. Copyright © 2014 Praise Worthy Prize S.r.l. - All rights reserved.

Keywords: Follower Density, Sequential Graph, Sequential Pattern Mining, Significance Measure, Spatio-Temporal Data Mining

temporal databases. Focus of the research today is on developing techniques to efficiently handle the voluminous spatiotemporal data. Considering the space and time characteristics of the events provides astuteness with regard to the evolutionary phenomena and influence of events on each other. It avails us comprehend the relationship between objects, events and their relationships. A spatio-temporal event is defined as an occurrence or happening in the real-world with a particular mention on location and time of occurrence. Events that have similar characteristics are collectively categorized to belong to the same event type. Spatio-temporal sequential pattern mining is defined as the process of extracting significant event type sequences from large spatio-temporal datasets. The work in this paper concentrates on extracting significant event type sequences from spatio-temporal event databases. We considered three major research demands in this paper: 1) definition of significance measure for discovering useful spatio-temporal sequential patterns, 2) data structure to represent the spatio-temporal database for reducing the run time of the sequential pattern mining process, 3) algorithm to mine spatio-temporal sequential patterns. The organization of the paper is as follows: In section 2, the problem statement is described. The review of related research is given in section 3 and section 4 defines the major contributions of this paper. The proposed grid-based algorithm for mining significant

Nomenclature STDB D(Gij) DGTh DF (ρ → f , Gij) DTh Ei ei FD (ρ → f ) FN (ρ → f ) G(x, y) Gij GR Nb(Gij) PN ( ρ → f ) SR SG ρ→f

Spatio-temporal database Density of grid cell Gij Minimum grid density threshold Density of follower node Gij of ρ → f Minimum density threshold An event type An event Follower density of sequential pattern ρ →f Follower nodes of sequential pattern ρ →f Size of grid cell Grid cell in ith row and jth column Spatial representation after applying grid Significant neighborhood of Gij Precursor nodes of sequential pattern ρ →f Spatial representation Sequential graph Sequential pattern with ρ as precursor and f as follower

I.

Introduction

Along with the ubiquitous computing devices, ubiquitous service infrastructure, context-awareness and open service architecture has evolved and gained importance, the area of spatio-temporal data mining. In scenarios such as moving objects, it can be said that perpetual amassment of spatial data prompts spatioManuscript received and revised March 2014, accepted April 2014


659

Gurram Sunitha, A. Rama Mohan Reddy

sequential patterns from the spatio-temporal event data set is described in detail in section 5. Section 7 provides the experimental results and comparative study of the proposed algorithm.

II.

Miner for large datasets, and it takes only one database scan to construct the slices. Unlike other approaches, which are based on spatial slicing, STS-Slicing-Miner used temporal slicing. The unidirectional property of time gives an advantage with temporal slicing that each slice can be discarded after being processed. [6] has discussed the complexity of analyzing the climate data due to its spatio-temporal nature. The work is concentrated on clustering and association analysis. A clustering methodology has been developed to find patterns that represent well-known climate indices. This methodology emphasizes on eliminating the limitations of traditional Eigen value analysis. Also, the experimental results of applying Apriori algorithm to extract the association rules from the time-series databases are evaluated. [7] has highlighted the limitations of traditional algorithms like GSP. A new tool CBSPM has been developed which can generate synthetic sequential datasets, and also handles user-specified constraints simultaneously. Mine fuzz change model, computes similarity computation index by using raw data collected from different time interval. This drastically increases the time complexity of the sequential pattern mining process. In [8], work has been done to eliminate the drawback of mine fuzz change model and an optimized fuzzy time interval algorithm is proposed. This algorithm extracts sequential patterns with less time complexity. [9] have made effort to mine frequent patterns of human interaction based on person-to-person meetings. Human interactions are modelled using a tree and the tree structure is analysed to mine interaction and their patterns that are frequent.

Problem Statement

In this section, we characterize the issue of mining spatio-temporal sequential patterns from spatio-temporal event data sets. Let STDB be the spatio-temporal database which consists of a set of event types {E1, E2, ...., En} where n represents the number of event types. Each of the event type Ei comprises of a set of events {e1, e2, ...., em}. Each event in the spatio-temporal database STDB consists of the fields identification of event (event ID), event type, spatio-temporal attributes and nonspatio-temporal attributes. We are considering solely the spatio-temporal attributes place of occurrence (location) and time of occurrence (time) of events to mine the sequential patterns. A spatio-temporal sequential pattern of form E1 → E2 → …. → Em can be defined as a significant event type sequence. The problem of spatio-temporal sequential pattern mining can now be defined as mining all significant sequential patterns from the given spatiotemporal database. In the process of mining sequential patterns from the spatio-temporal database, the first challenge is to reduce the number of database scans needed to generate sequential patterns and the second challenge is to design a significance measure to prove consequentiality of the sequential patterns.

III. Related Research: a Brief Review For effectual mining of sequential patterns from spatio-temporal databases, the literature presents a handful of researches. Since approaches for spatiotemporal sequential pattern mining has picked up immense paramountcy in real-life applications lately, a succinct survey of the crucial researches cognate to the same is presented here. [1] has introduced the idea of sequential pattern mining and has confirmed its momentousness in the field of data mining. The sequential patterns are extracted from the market-basket data and are used to predict customer buying patterns. Since the introduction of sequential pattern mining, many algorithms have been proposed that efficiently work on sequence databases [2][4]. [5] investigated the problem of finding spatiotemporal sequential patterns, from a large database of spatio-temporal event data sets. To identify significant sequential patterns, sequence index has been defined as the significance measure. Two algorithms were discussed. A simple algorithm called STS-Miner that mines all the significant event type sequences is illustrated. STS-Miner algorithm works naive with small databases that can fit entirely in the main memory, whereas it becomes inefficient in case of large databases. Slicing-STS-Miner is proven to be faster than STS-

IV.

Contribution of the Paper

In this paper, we focus on tending to the two issues specified in section 2 related to mining the sequential patterns from spatio-temporal databases. The contribution of this paper is given below:  Sequential graph: Sequential Graph is a data structure containing information about significant grid cells called as principle nodes. Once the sequential graph is constructed from the spatiotemporal database, all the sequential patterns can be derived from the sequential graph itself. This reduces the number of database scans as well as the run time of the data mining algorithm.  Significance measure: To mine significant sequential patterns from the spatio-temporal database, we designed a proficient significance measure called as follower density.

V.

Developing a Grid-Based Algorithm for Mining Sequential Patterns

Extracting the set of significant sequential patterns from the spatio-temporal database is the eventual goal.



660


Illustrated below is the anticipated sequential patternmining algorithm, which is split into two major steps:  Preprocessing Phase.  Mining Phase. Fig. 1 illustrates the block diagram of the anticipated grid-based sequential pattern mining algorithm. In the preprocessing phase, the first step is to build spatial representation of the spatio-temporal database and to apply rectangular grid on the spatial representation. Consequently, all the events in the database will be grouped into grid cells. Then for each grid cell, its density is calculated and the cells with insignificant density are eliminated. The idea is to eliminate sparse grid cells. The cells with density greater than the minimum grid density threshold are considered to be the principle nodes. In the second step which is mining phase, sequential graph is constructed with the help of the principle nodes and then sequential patterns are mined from the sequential graph, based on the defined significance measure.

As spatio-temporal databases are large, number of scans of the database is the major factor that influences the run time of the algorithms. With the intention of reducing the number of scans of the database, we have proposed a data structure called as sequential graph. As principle nodes only are considered for constructing the sequential graph, the final graph will be smaller in size. Implicitly, we are reducing the database size with the help of minimum grid density threshold. V.2.1.

Identifying Principle Nodes

Fig. 3 shows the spatial representation SR of the database shown in Fig. 2, with x-axis and y-axis representing linear space and time respectively. The spatial representation SR is then mapped onto a rectangular grid where grid cell size is user defined. Let G(x, y) define the size of the grid cell, where x defines spatial neighborhood and y defines temporal neighborhood.

Fig. 1. Block diagram of the proposed algorithm

V.1.

Spatio-Temporal Database

Fig. 3. Spatial representation of the spatio-temporal database

Fig. 2 represents the example spatio-temporal database STDB. It contains 4 event types A, B, C and D. Each event type comprises of a set of events where each event is represented as event ID (location, time).

Fig. 4 shows GR , the spatial representation after constructing the grid, with x =1 and y = 1. Each grid cell is given a numeric code which serves as identification for the cell. Gij is the grid cell where i and j determine the location of the cell on x-axis and y-axis respectively.

Fig. 4. Spatial representation after constructing grid

Fig. 2. A Spatio-Temporal Database

V.2.

Sequential graph SG is constructed by adding up grid cells as nodes to the graph. Yet, including all the cells to the sequential graph leads to large graph and hence to more mining time.

Pre-Processing Phase

To assemble the sequences of data, normally many of the sequential pattern mining algorithms involve a number of database scans. Copyright © 2014 Praise Worthy Prize S.r.l. - All rights reserved


661


Hence, considering the fact that not all grid cells will produce significant sequential patterns, eliminating cells with insignificant grid density will reduce the number of cells added to the graph. In this regard, we define the density of the grid cell.

But, only {G22, G32} can be considered as significant neighborhood nodes because only these are principle nodes. Table I shows the list of principle nodes and their corresponding significant neighborhood nodes. Table II shows the principle nodes and their comprising events. The comprising events may be of one or more event types. Examine the node G21 which has two events B11 and B42 where 11 and 42 represents the identification of the event (e11 and e42) and the symbol ‘B’ identifies the event type.

Definition 1. Density of a Grid Cell Gij: is the ratio of number of events in the grid cell Gij to the total number of events in STDB. The following equation defines the density of the grid cell Gij:

D  Gij  

e | e  Gij e | e  STDB

(1)

TABLE I PRINCIPLE NODES AND THEIR CORRESPONDING SIGNIFICANT NEIGHBORHOOD NODES Principle Significant Principle Significant Node Neighborhood Node Neighborhood G21 { G22 , G32} G33 { G34 , G43} G22 { G32 , G33} G43 { G33 , G34 , G53} G32 { G22 , G33 , G42 , G43} G53 {G43} { G32 , G33 , G43 , G52 , G42 G34 { G35} G53} G52 { G42 , G43 , G53} G35 --

where e is an event in STDB. After constructing the grid, the density of each grid cell is calculated and the cell is considered to be significant only if its density D(Gij) is greater the minimum grid density threshold value DGTh as given in the equation below:

D  Gij   DGTh

(2)

The cells that are significant are called as principle nodes. Only principle nodes are used to construct the sequential graph SG. The minimum grid density threshold value DGTh is user-defined.

TABLE II PRINCIPLE NODES AND THEIR COMPRISING EVENTS Principle Node

Example: After calculating the density of each grid cell in Fig. 4 using equation (1) and considering the minimum grid density threshold value as 0.03, the cells G21,G22, G32, G33, G34, G35, G42, G43, G52 and G53 are found to be the principle nodes.

Comprising Events

{ B11 , B42}

G33

G22 G32

{ A12 , B19 , B25 , B32 } { B06 , B07 ,B08 , B21, C10 } { B09 , D16 } { A03 , A04 , C26, C37 , C40 }

G43 G53

{ A23 , B02 , B24 , B38 , C44 } { A17 , A22 , B14 } {A30 , C15 , C29 , C41 }

G34 G35

{ C28 , D27 , D35 , D39} { C34 , D36 , D43 }

Construction of Sequential Graph

Once the principle nodes are identified, to construct the sequential graph, the neighborhood of the set of principle nodes needs to be examined. With the intention of finding the neighborhood nodes of a principle node Gij, we consider the nodes on both sides of Gij on x-axis (spatial neighborhood) and only the upper side nodes on y-axis (temporal neighborhood) due to the unidirectional property of time. Based on this reasoning, the maximum number of neighborhood nodes for a node is five.

Principle Node

G21

G42 G52

V.2.2.

Comprising Events

The next step is to assemble the sequential graph with the aid of the principle nodes, their significant neighborhood nodes and comprising events. Principle nodes form the vertices of the sequential graph with edges connecting each principle node with its significant neighborhood nodes. The first node from the set of principle nodes is considered as root node. Each node of the sequential graph contains the information - identification of the principle node and its comprising events. The sequential graph built based on the information from Tables I and II is shown in the Fig. 5. Algorithm for the pre-processing phase is shown in Fig. 6.

Definition 2. Significant Neighborhood Nb(Gij) : For a node Gij, the neighborhood nodes are defined as {Gi-1, j ,Gi+1, j , Gi-1, j+1 , G i, j+1 , Gi+1, j+1 } where i, j represent the location of the grid on the x-axis and y-axis respectively. For a principle node Gij, among its five neighborhood nodes only principle nodes are considered as the significant neighborhood Nb(Gij) i.e., the significant neighborhood nodes of a principle node must also be principle nodes.

V.3.

Mining Spatio-Temporal Sequential Patterns from the Sequential Graph

Once the sequential graph is constructed, sequential patterns can be generated using sequential graph without the need to scan the spatio-temporal database.

Example: Consider the principle node G21, which has the set of neighborhood nodes as {G11, G31, G12, G22, G32}. Copyright © 2014 Praise Worthy Prize S.r.l. - All rights reserved


662


sequential pattern ρ → f, the precursor nodes are the set of nodes in the sequential graph that contain at least one event of event type ρ. Definition 5. Follower Nodes FN (ρ → f ): For a sequential pattern ρ → f, the follower nodes are the set of nodes in the sequential graph containing at least one event of event type f in the neighborhood nodes of the precursor nodes of ρ → f i.e., in {Nb ( PN ( ρ → f ) ) }. The definition can be recursively applied for sequential patterns of length ≥ 3. Consider a sequential pattern ρ1 → ρ2 → …. ρk → f of length k+1, where k ≥ 2. The follower nodes of sequential pattern ρ1 → ρ2 → …. ρk becomes the precursor nodes for sequential pattern ρ1 → ρ2 → …. ρk → f . Fig. 5. Sequential Graph of the spatio-temporal database

Definition 6. Density of a Follower Node of ρ → f : Let Gij be one of the follower nodes of sequential pattern ρ → f. The density of the follower node Gij for sequential pattern ρ → f is defined as the ratio of number of events of event type f in Gij to the total number of events in Gij.

Algorithm 1 Pre-processing Input : Spatio-Temporal Dataset STDB Output : Sequential Graph SG = {V, E} where V – set of vertices E – set of edges Variables Pl : Set of Principle Nodes N : Significant Neighborhood





DF   f , Gij 

e | e  Gij  e  f  e | e  Gij

(3)

1: SR  STDB where e is an event.

G x, y  2: GR   S R 3: P   4: for each grid cell Gij in GR do 5: Calculate density D(Gij) (using equation 1) 6: if D(Gij) ≥ DGTh then //Gij is principle node 7: P  P  Gij 8: end if 9: end for 10: V  P //principle nodes form vertices of SG 11: E   12: for each principle node Gij in Pl do 13: N  NbGij 14: for each node Cij in N do E  E  Gij , Cij // add edge from Gij to Cij 15: 16: end for 17: end for

Definition 7. Follower Density FD ( ρ → f ): Follower Density defines the significance of a sequential pattern and is used to eliminate spurious patterns. For a sequential pattern ρ → f , the follower density is defined as the average of the sum of the densities of the follower nodes of ρ → f :

FD    f  

1 F

 DF    f ,G 

(4)

GF

where F is the set of Follower Nodes of ρ → f i.e., F = {FN (ρ → f)}. The mining procedure starts off by considering the existing event types in the sequential graph as the set of precursors. Subsequently, sequential patterns are generated by selecting followers for each precursor. Consider ρ as one of the precursors. The first step is to discover the precursor nodes for ρ. These are the nodes in the sequential graph which contain at least one event of event type ρ. Then the neighborhood nodes of the precursor nodes are found and all the existing event types in them are listed out. These form the set of follower event types for the precursor ρ. Once the set of follower event types are determined, potential sequences can be generated. Let { f1 , f2 ,…., fk }be the set of follower event types for precursor ρ.

Fig. 6. Algorithm for Pre-processing phase

Definition 3. Precursor ρ and Follower f: Consider a spatio-temporal sequential pattern ρ → f of length 2 where ρ and f are event types. Here ρ is said to be the Precursor and f is said to be the Follower. The sequential pattern says that Follower f follows Precursor ρ. Consider a sequential pattern ρ1 → ρ2 → …. ρk → f of length k+1, where k ≥ 2. Here ρ1 → ρ2 → …. ρk is said to be the Precursor and f is said to be the Follower. Definition 4. Precursor Nodes PN (ρ → f ): For a



663


The potential sequences would be { ρ → f1 , ρ → f2 , ….. , ρ → fk }. For each sequence, then find the follower density. The sequence ρ → f is considered as significant sequential pattern if FD (ρ → f) ≥ DTh ,where DTh is the minimum follower density threshold defined by the user. Similar procedure can be used to estimate the follower density of sequences with length ≥ 3. Grid-based Algorithm for mining significant sequential patterns from a given spatio-temporal database is shown in Fig. 7.

Algorithm 2 GMine Input : SG - Sequential Graph Output : FSEQ - Set of Significant Sequential Patterns Variables ET (Ni) - Set of Event Types in Node Ni NEi - Set of Precursor Nodes maxLen - Maximum Length of sequences to be generated

Example: From the Sequential graph shown in Fig. 5, the existing event types are {A, B, C, D}. These form the set of precursors from which we select one event type at a time as precursor and subsequently produce the sequences by adding follower event type to the precursor. For example, consider A as the precursor. The precursor nodes for A are determined as {G22, G33, G43, G52, G53}. Consequently their neighborhood nodes are {G32, G33, G34, G42, G43, G53} and the set of follower event types are {A, B, C, D}. Hence, the set of potential sequences are {A→A, A→B, A→C, A→D}. Now for calculating the follower density of the potential sequence A→B, the follower nodes are determined as {G32, G33, G42, G43}. The follower density FD(A→B) is calculated using Eq. (2): 1 4 3 1 1 FD(A→B) =      = 0.558 4  5 5 2 3

E for each node Ni in SG do E  E  ET Ni  end for FSEQ   for each event type Ei in E do NEi  PN Ei 

8: 9:

GenSeq ( Ei , 1 , maxLen , NEi ) end for

Algorithm 3 GenSeq Input: ρ – Sequence to be expanded L – Length of ρ maxLen - Maximum Length of sequences P – Set of Precursor Nodes of ρ Variables N – Set of Neighborhood Nodes E - Set of Follower Event Types newSeq – New Sequence FnewSeq – Set of Follower Nodes 1: N   2: for each Precursor Node Pi in P do N  N  Nb Pi  3: 4: end for 5: E   6: for each Neighborhood Node Ni in N do 7: E  E  ET Ni  8: end for 9: for each Follower Event Type Ei in E do newSeq  ρ  Ei  10: FnewSeq  FNnewSeq  11: 1 12: FDnewSeq    DFnewSeq, G  FnewSeq GFnewSeq

Consider a sequence A→B→C of length 3, where A→B is precursor and C is follower. The follower nodes of A→B will become the precursor nodes for A→B→C. Hence, the precursor nodes are {G32, G33, G42, G43} and the follower nodes are {G32, G33, G34, G52, G53}. The follower density of the sequence FD(A→B→C) is calculated as:

11 1 1 3 3 FD  A  B  C          0.4 55 5 4 5 4

VI.

1: 2: 3: 4: 5: 6: 7:

Results and Discussion

The investigational results of the proposed grid-based sequential pattern mining algorithm is described in this section. A comparative study of the proposed algorithm with the STS-miner algorithm [5] is done using the synthetic datasets and the real world datasets.

13: 14: 15: 16: 17: 18: 19:

VI.1. Datasets The proposed approach is implemented using Java. The experimentation has been carried out on a PC machine with Pentium dual core 3Mhz processor and 2 GB main memory.

if FDnewSeq   DTh then FSEQ  FSEQ  newSeq if L+1 < maxLen then GenSeq( newSeq , L+1 , maxLen , FnewSeq ) end if end if end for Fig. 7 Grid-based algorithm for mining sequential patterns



664


Dataset 1: (synthetic data): With space and time values, we have generated the synthetic data that comprise of four event types and 1000 events. Dataset 2: (Real world data): From the UCI machinelearning repository, we have taken the real world data, ‘Localization Data for Person Activity Data Set’ [14]. This dataset comprises of 11 activities (event types) and 164860 events. VI.2. Results and Discussion Fig. 8 and Fig. 9 show the performance comparison of the two algorithms with respect to the length of the sequential patterns. Dataset 1 is used for the purpose. For conducting the experiment using STS-miner algorithm, length of sequential patterns is varied and respective run time for generating the patterns is recorded. Other parameters for STS-miner algorithm are set as – R = 1, T = 1, SI = 0. For grid-based algorithm parameters are set as – G(x, y) = (1, 1), DGTh = 0, DTh = 0. The results in Fig. 8 show that Grid-based algorithm takes less time to generate the sequential patterns of the given length when compared to STS-miner. Fig. 9 shows the memory usage of the algorithms with respect to the length of the sequential patterns. It is seen that the memory usage of the grid-based algorithm remains constant for the given spatio-temporal database, G(x, y), DGTh and DTh. Fig. 10 and Fig. 11 show the performance of the grid-based algorithm when run on Dataset 2. Fig. 10 shows that the size of the sequential graph grows linearly with the database size. The remaining parameters are fixed as - G(x, y) = (1, 1), DGTh = 0 and DTh = 0. Fig. 11 shows that the run time of the grid-based algorithm decreases with the increase in the minimum grid density threshold, DGTh. The remaining parameters are set as - G(x, y) = (0.5, 0.5) and DTh = 0.1 and the length of sequential pattern is taken as 10 for the experimental purpose.

Fig. 9. Memory Usage with respect to length of sequential patterns

Fig. 10. Memory Usage of grid-based algorithm with respect to the spatio-temporal database size

Fig. 11. Run Time Performance of grid-based algorithm with respect to DGTh

VII. Fig. 8. Run Time Performance with respect to length of sequential patterns

Conclusion

We have presented an efficient approach for mining



665


[13] Dataset from UCI machine learning repository, http://archive.ics.uci.edu/ml/datasets/Localization+Data+for+Pers on+Activity

spatio-temporal sequential patterns. We have produced solutions for three major challenges: (1) Definition of significance measure for finding useful spatio-temporal sequential patterns, (2) Data structure to represent the database for reducing the run time of the mining process, (3) Algorithm to mine the spatio-temporal sequential patterns. The experimentation on synthetic and real world datasets ensured the effectiveness of the proposed approach in terms of self-regulating the significant sequential patterns and computation time compared with the STS miner. Further research may include analysis of the effects of grid size, minimum grid density threshold and minimum density threshold for identifying significant sequential patterns. Alternate neighborhood functions and data structures may be defined and worked upon. Further, the grid-based algorithm may be customized to suit specific applications.

Authors’ information Gurram Sunitha is an Associate Professor in the Department of CSE at Narayana Engineering College, Nellore, A.P., India. She has 14 years of teaching experience at both UG and PG levels. She has completed B.E. in Electronics & Communications Engineering from Gulbarga University in 1999 and M.Tech in Computer Sciences from JNT University, Anantapur in 2005. Currently she is pursuing Ph.D. in Computer Science and Engineering at S.V.University, Tirupati. Her research interests include Data Mining, Automata Theory and Database Systems. Dr. A. Rama Mohan Reddy received his B.Tech. degree from JNT University, Anantapur in 1986, M. Tech degree in Computer Science from NIT, Warangal in 2000 and Ph.D. in Computer Science and Engineering in 2008 from S. V. University, Tirupathi. He is presently working as Professor in Department of Computer Science and Engineering, S. V. University College of Engineering, Tirupathi, A.P. India. His research interests include Software Architecture, Software Engineering and Data Mining. He is life member of ISTE and IE.

References [1]

R. Agrawal and R. Srikant, “Mining Sequential Patterns”, Proc. 1995 Int’l Conf. on Data Eng. (ICDE ‘95), Mar. 1995. [2] J. Han, J. Pei, B. Mortazavi-Asl, Q. Chen, U. Dayal, and M.-C. Hsu, “Freespan: Frequent Pattern-Projected Sequential Pattern Mining”, Proc. 2000 ACM SIGKDD Int’l Conf. Knowledge Discovery in Databases (KDD’00), Aug. 2000. [3] M. J. Zaki, “SPADE: An Efficient Algorithm for Mining Frequent Sequences”, Machine Learning, Vol. 42, 2001. [4] J. Pei, J. Han, B. Mortazavi-Asl, J. Wang, H. Pinto, Q. Chen, U. Dayal, and M.-C. Hsu, “Mining Sequential Patterns by PatternGrowth: The PrefixSpan Approach”, IEEE Trans. Knowledge and Data Eng., Vol. 16, No. 11, Nov. 2004. [5] Yan Huang, Liqin Zhang, and Pusheng Zhang, "A Framework for Mining Sequential Patterns from Spatio-Temporal Event Data Sets", IEEE Transactions On Knowledge And Data Engineering, Vol. 20, No. 4, Apr. 2008. [6] P. Zhang, M. Steinbach, V. Kumar, S. Shekhar, P. Tan, S. Klooster and C. Potter, “Discovery of Patterns of Earth Science Data Using Data Mining”, Next Generation of Data Mining Applications, 2004. [7] Yilmaz, G., Badur, B.Y., Mardikyan, S., Development of a constraint based sequential pattern mining tool, (2011) International Review on Computers and Software (IRECOS), 6 (2), pp. 191-198. [8] Mary Gladence, L., Ravi, T., Mining the change of customer behavior in fuzzy time-interval sequential patterns with aid of Similarity Computation Index (SCI) and Genetic Algorithm (GA), (2013) International Review on Computers and Software (IRECOS), 8 (11), pp. 2552-2561. [9] Uma, S., Suguna, J., Tree-based weighted interesting pattern mining approach for human interaction pattern discovery, (2013) International Review on Computers and Software (IRECOS), 8 (11), pp. 2570-2575. [10] H. Cao, N. Mamoulis, and D.W. Cheung, “Mining Frequent Spatio-Temporal Sequential Patterns”, Proc. Fifth IEEE Int’l Conf. Data Mining (ICDM ’05), 2005. [11] R. Srikant and R. Agrawal, “Mining Sequential Patterns: Generalizations and Performance Improvements”, Proc. Fifth Int’l Conf. Extending Database Technology (EDBT ’96), Mar . 1996. [12] Hugo Alatrista Salas, Sandra Bringay, Frédéric Flouvat, Nazha Selmaoui-Folcher and MaguelonneTeisseire, "The Pattern Next Door: Towards Spatio-sequential Pattern Discovery", Advances in Knowledge Discovery and Data Mining, Lecture Notes in Computer Science, Vol. 7302, 2012.



666


Online Modules Placement Algorithm on Partially Reconfigurable Device for Area Optimization Mehdi Jemai, Bouraoui Ouni, Abdellatif Mtibaa Abstract – In this paper we propose a new method of modules placement on the partially reconfigurable device. The main aim of our method is to optimize the area occupation of the device when placing the modules. The key word of our algorithm is the not consideration of the heights and the widths of modules, like former algorithms. In fact, our algorithm uses the size of modules; next it computes the heights and the widths of modules, and after that it places them on the device while optimization the area occupation. Moreover, most of existing algorithms do not evaluate their algorithms on real hardware device; we do experimentations of our tasks on a real FPGA Virtex-5. Results show that our algorithm achieves better placement quality in term of area occupation and faster in term of run time compared to existing approaches. Copyright © 2014 Praise Worthy Prize S.r.l. - All rights reserved.

Keywords: Online Module Placement Algorithm, Partially Reconfigurable Device, FPGA

I.

This algorithm adds the alignment flag of each free segment to determine the placement location of the task with the corresponding free segment. In [8] Lu et al. proposed their 1D Reuse and Partial Reuse (RPR) algorithm for to reduce the reconfiguration time by reusing already placed tasks. In [9] Zhou et al. proposed their 2D Window-based Stuffing algorithm to improve the 2D stuffing by using time windows instead of the time events, but her disadvantage is the long execution time. In [10] the authors proposed a new algorithm to reduce the runtime cost, the proposed approach computed of the earliest available time (EA) matrix for every incoming task; that contains the earliest starting times for scheduling and placement of the arriving task. In [11] and [12] T. Marconi et al. proposed two algorithms the Intelligent Merging (IM) and the Quad-Corner (QC). The IM algorithm combines three techniques: 1) The Merging Only if Needed (MON) technique which allows to merge blocks of empty area only if there is no available block for the incoming task and to terminate merging process earlier. 2) The Partial Merging (PM) technique gives an ability to merge only a subset of the available blocks so as To reduce the algorithm execution time. 3) The Direct Combine (DC) technique is combines directly the blocks to increase the placement quality. The IM algorithm uses with these techniques the Combine Before Placing (CBP) strategy that allows directly merges blocks to form a bigger block before placing a task when possible. The QC algorithm spreads the arriving hardware tasks to the four corners of the devices instead the one corner to solve splitting free area problem. In [13] authors proposed a placement technique allow reducing the reconfiguration time.

Introduction

Partial Reconfiguration (PR) is the ability to reconfigure part of FPGA devices while the rest of the device still running. This approach is useful for systems with multiple functions that can time-share the same device resources. In such systems, one section of the FPGA continues to operate, while other sections of the FPGA are reconfigured to provide new functionality. Hence, designer can increase the functionality of the FPGA; because the PR improves logic density by removing the need to implement functions that do not operate simultaneously in the FPGA. However, one challenge to be faced when using PR is the scheduling and placement of modules. This problem has been called temporal placement [1], [2], [3]. In [1] the authors have proposed a typical model to solve the temporal placement problem. The model is based on a scheduler and a placer. The scheduler calculates the time at which a task should be executed and the placer places the selected task into available reconfigurable hardware resources. Many algorithms have been proposed to solve the scheduling and placement issue, such as: Horizon [4] and Stuffing [5] proposed by Steiger et al. for 1D and 2D area models. The Horizon guarantees that arriving tasks are only scheduled when they do not overlap in time or space with other scheduled tasks. Then, the Stuffing schedules arriving tasks to arbitrary free areas that will exist in the future by imitating future task terminations and starts. In [6] Chen and Hsiung proposed their 1D Classified Stuffing algorithm which will be classified incoming tasks before scheduling and placement. In [7] T. Marconi et al. proposed their 1D Intelligent Stuffing to solve the problems of both the 1D Stuffing and Classified Stuffing.



667

Mehdi Jemai, Bouraoui Ouni, Abdellatif Mtibaa

In [14] authors combined temporal partitioning and temporal placement to reduce the communication cost between modules. Now, in this paragraph, we interpret the above algorithms. Most of existing algorithms have considered the modules as rectangular shape with predefined dimension. The dimension is characterized by two parameters W (width) is the number of columns and H (height) is the number of lines. Next the author proposed a placement algorithm that aims to optimize one or more design parameters such as the area overhead, the latency, the reconfiguration time, etc, while respecting all constraints. Hence, most of former algorithms can be formulate as follows:  Given a set of modules to be placed, each module is characterized by its (W, H), a reconfigurable device and a set of constraints. Find a possible placement of modules on the device that optimizes one (or more) design parameters and while respecting all constraints. However, in this paper we propose a new modules placement methodology that optimizes the area occupation when placing modules on the device. The main key words of our methodology is the following: Our methodology does not consider the widths and the heights of modules in fact it is based on the area of each module, next it calculates its width and the height. Hence, our method, see Fig. 1, can be formulated as follows:  Given a set of modules to be placed, each module is characterized by its area (not its W, H), a reconfigurable device and a set of constraints (see Fig. 1), find: 1) the W and H of each module. 2) The coordinates (xi , yi) of each module on the device that optimizes the area occupation and respecting all constraints.

Fig. 2. The holes

Fig. 3. Flowchart of our proposed algorithm

Fig. 1. Proposed algorithm

II.

Proposed Algorithm

When placing modules on the device many holes, unallocated areas, may be created as shown in Fig. 2. Our algorithm aims to optimize the area occupation of the device when placing these modules. This can be reached by finding the adequate placement of the modules that minimize number of holes on the device. Our algorithm is based on Keeping Non-overlapping Empty Rectangle (KNER) method [15]. Our algorithm follows the flowchart given by the Fig. 3. The pseudo-code of our algorithm is shown in Fig. 4.

Fig. 4. Pseudo-code of our algorithm



668


The functions “verify" return the height [i] and the width [i] of the module [i]. The following pseudo-code describes the function "verify" (Fig. 5).

Fig. 5. Pseudo-code of the function "verify"

II.1.

Illustrative Example

Applying the steps of our algorithm on the modules of Table I, the size of chip has been assumed 40×50 a 2dimensional CLB array. TABLE I THE MODULES Modules M1 M2 M3 Size(CLB) 80 24 130

Figs. 6. Placement of modules

III.1. Simulation Experiment In this part, we have conducted two kinds of experiments. In the first experiment we have fixed the order of modules and we have changed the size of the device. In the second one, we have fixed the size of the device and we have changed the order of the modules. We have compared our algorithm to Bazargan’s algorithms like First Fit algorithm (FF), Best Fit algorithm (BF), Bottom Left algorithm (BL) and Least Interference Fit algorithm (LIF) [17]; also with QuadCorner algorithm (QC) [12].

M4 42

The first module M1 has as size of 80 CLB, we will apply our algorithm to determine the height and the width of this module as follows: w1=4

(1)

a = M1 mod w1 = 80 mod 4 = 0

(2)

h1 = M1 div w1= 80 div 4 = 20

(3)

III.1.1. First Experiment In this experiment, we consider the following dimensions of the chip 80x120, 120x120 and 120x80 a 2dimensional CLB array. We have used sixty modules that arrive in order, one after one, from the first module to the last one. In this experiment we have considered the wasted area and run time of the algorithm as design metrics to evaluate the design results. Figs. 7 give a scream capture, in which we show how each algorithm places the modules on the device. Figs. 7 show that our algorithm gives 1 hole, BL 11 holes, FF 16 holes, BF 13 holes, LIF 14 holes and the QC 9 holes. Hence, our algorithm provides an improvement of 90,90% compared to BL, 93,75% compared to FF, 92,30% compared to BF, 92,85% compared to LIF and 88,88% compared to QC. Fig. 8 gives the percentage of wasted that given by each algorithm. The percentage of wasted area has been calculated as follows:

Therefore, the height and the width of the first task are respectively 20 and 4. It placed in the coordinate (0; 0) as shown in Fig. 6(b). Similarly we determined the height and the width of each module: M2 (6, 4), M3 (33, 4) and M4 (11, 4). To implement each module, we must place it in the first suitable NER starting from left to right. Instantly we have the following NER: (abcd) and (befg). We note that the NER (abcd) is suitable for the module M2 as shown in Fig. 6(c). By the same way our algorithm places the rest of modules (M3, M4) on the device, as shown in Figs. 6.

III. Experiments and Discussion To fairly evaluate our proposed algorithm, we have done tow kind of experiments: simulation and physic. For simulation experiment, we have built a framework in JAVA language. The framework run under Windows-7 on a Intel Core 2 Duo T5500, 1,66 GHz, 1GB of RAM. Next, we have re-implemented physically and partially the modules on real Xilinx FPGA Virtex-5.

Area of the holes × 100 Area of the device

(4)

Fig. 9 shows the average wasted area given by each algorithm.



669


Fig. 11 shows the average free area of each algorithm, this figure shows that the average free area of our algorithm is 5907 CLB, the FF is 5800 CLB, BF is 5842 CLB, BL is 5815 CLB, LIF is 5820 CLB and QC is 5861 CLB. Hence, the gain in area of our algorithm is 94,69% compared to FF, 92,30% compared to BF, 93,93% compared to BL, 93,54% compared to LIF and 88,46% compared to QC algorithm. Fig. 12 gives the run time of each algorithm in millisecond. Fig. 13 shows the average run of time of each algorithm, this figure shows that the average run of time of FF is 18,18 ms, BF is 20,15 ms, BL is 17,76 ms, LIF is 19,17 ms, QC is 21,30 ms and our algorithm is 9,48 ms. Hence, result show that our algorithm is faster 1,91 time compared to FF, 2,12 time compared to BF, 1,87 time compared to BL, 2,02 time compared to LIF and 2,24 time compared to QC algorithm.

Figs. 7. The scream capture

Fig. 10. The free area

Fig. 8. Wasted area

Fig. 11. The average of free area

Fig. 9. The average of wasted area

Fig. 9 shows that the average of wasted area obtained by the FF is 1,04%, BL is 0,93% , BF is 0,67%, LIF is 0,84% , QC is 0,48% and our algorithm is 0,06%. Fig. 10 gives the free area given by each algorithm. The free area has been calculated as follows: Fig. 12. Run time

Area of the device – (Wasted area + area of modules) (5)



670


time of FF is 31,99 ms, BF is 34,53 ms, BL is 29,85 ms, LIF is 33,56 ms, QC is 28,38 ms and our algorithm is 9,47 ms.

Fig. 13. The average run of time

III.1.2. Second Experiment In the second experiment, we have fixed the size of the device to 120×120 and we have we have placed randomly the modules on the devices. We have repeated this experiment three times. Fig. 14 gives the % of wasted area provided by each algorithm. Fig. 15 show that the average of wasted area given by each algorithm. Fig. 15 shows that the average of wasted area obtained by the FF is 0,62%, BF is 0,61% , BL is 0,64%, LIF is 0,64%, QC is 0,33% and our algorithm is 0,03%. Hence, the gain in area of our algorithm is 94,44% compared to FF, 94,31% compared to BF, 94,56% compared to BL, 94,62% compared to LIF and 89,58% compared to QC algorithm.

Fig. 16. Run time

Fig. 17. The average run of time

Hence, result show that our algorithm is faster 3,37 time compared to FF, 3,64 time compared to BF, 3,15 time compared to BL, 3,54 time compared to LIF and 2,99 time compared to QC algorithm. III.2. Real Experiment To confirm our approach, in this part of experiment, we have physically re-implemented the modules shown in Figs. 6 on FPGA Xilinx Virtex-5 XC5VLX50T. To meet this aim we have used Xilinx ISE tool, Xilinx PlanAhead tool and the Early Access approach [18]. Design results are calculated by Xilinx-ISE CAD tool that provides resources and timing report incorporates timing delay and resources to provide a comprehensive area and timing summary of the design. Design results are shown in Table II, result show that our algorithm provides a gain in area of 11, 85 % compared to FF algorithm,8,8% compared to BF algorithm, 7,3 % compared to BL algorithm, 11,55% compared LIF algorithm and 9,40% compared to QC algorithm. Also, results show that our algorithm provides a gain in whole time by 34,12% compared to FF algorithm, 33,49% compared to BF algorithm, 31,63 % compared to BL algorithm, 44,88% compared LIF algorithm and 41,79 % compared to QC algorithm:

Fig. 14. Wasted area

Fig. 15. The average of wasted area

Fig. 16 gives the run time of each algorithm in millisecond. Fig. 17 shows the average run of time of each algorithm, this figure shows that the average run of



671


pp. 132-138, 2007. [11] T. Marconi, Y. Lu, K.L.M. Bertels, G.N. Gaydadjiev, "Intelligent Merging Online Task Placement Algorithm for Partial Reconfigurable Systems", Proceedings of Design, Automation and Test in Europe (DATE), March 2008. [12] T. Marconi, Y. Lu, K.L.M. Bertels, G.N. Gaydadjiev, "A Novel Fast Online Placement Algorithm on 2D Partially Reconfigurable Devices", Proceedings of the International Conference on FieldProgrammable Technology (FPT), December 2009 [13] B Ouni, A Mtibaa, "Optimal placement of modules on partially reconfigurable device for reconfiguration time improvement", Microelectronics International, Vol. 29 Iss: 2, pp.101-107, 2012 [14] B Ouni, R Ayadi, A Mtibaa, "Combining Temporal Partitioning and Temporal Placement Techniques for Communication Cost Improvement", Advances in Engineering Software, Elsevier Publishers. (ISSN: 0965-9978), Volume 42 (Issue 7), pp 444-451, 2011 [15] T Marconi, "Efficient Runtime Management of Reconfigurable Hardware Resources", Delft University of Technology, pp 13-17, June 29, 2011 [16] Xilinx Inc. Two flows for partial reconfiguration: module based or difference based. Xilinx Application Note XAPP290 Sep, 2004. [17] K. Bazargan, R. Kastner, and M. Sarrafzadeh. Fast Template Placement for Reconfigurable Computing Systems. In IEEE Design and Test of Computers, vol. 17, pp. 68-83, 2000. [18] Xilinx, Early Access Partial Reconfiguration User Guide UG208 (v1.1) March 6, 2006.

Whole time= Route time+ Place time+ Mapping time (6)

Algorithm area occupation% Route time (s) Place time (s) Mapping time (s) Whole time (s)

TABLE II THE DESIGN RESULTS Prop. FF BF algo 20 22,37 21,76 59 80 75 86 145 131 110 165 178 255 390 384

IV.

BL

LIF

QC

21,46 72 127 168 367

22,31 72 177 224 473

21,89 48 160 225 433

Conclusion

In this paper, we have proposed behavioral algorithms useful for reconfigurable computing systems. The proposed algorithms can be used early, at behavioral level, in the design flow and it is used to reduce the wasted area on the device when placing modules on it. The proposed algorithms have been tested and compared to others algorithms used in this field. After that, we have refined the design down to the implementation levels. Physical results show that our algorithm keeps its assets at lower design levels.

Authors’ information

References

Mehdi Jemai held a Diploma in Computer Engineering in 2009 from the Higher Institute of Applied Science and Technology of Sousse and received his Master in Microelectronic in 2011 from the Faculty of Science of Monastir. Currently, he prepares, in the Engineering School of Monastir, his thesis whose interest includes methodologies development for reconfigurable architectures.

[1]

A. Mtibaa, B. Ouni, M. Abid, "An efficient list scheduling algorithm for time placement problem", Computers Electrical Engineering, Volume 33 (Issue 4), pp 285-298, 2007. [2] Bobda Ch, "Introduction to Reconfigurable Computing Architectures, Algorithms, and Applications", Springer Netherlands, ISBN 978-1-4020-6088-5, 5 (Print) 978-1-40206100-4 359, 2007. [3] P Mahr, S Christgau, C Haubelt, C Bobda, "Integrated Temporal Planning, Module Selection and Placement of Tasks for Dynamic Networks-on-Chip", Proceeding of IEEE International Symposium on Parallel and Distributed Processing Workshops and PhD, pp 258-263, 2011. [4] C. Steiger, H. Walder, and M. Platzner, "Heuristics for Online Scheduling Real-Time Tasks to Partially Reconfigurable Devices", Proceeding of Field-Programmable Logic and Applications (FPL), LNCS 2778, pp. 575-584, 2003. [5] C. Steiger, H. Walder, and M. Platzner, "Operating Systems for Reconfigurable Embedded Platforms: Online Scheduling of RealTime Tasks", IEEE transaction on Computers, Vol. 53, No. 11, pp. 1393-1407, 2004. [6] Y. Chen and P. Hsiung, "Hardware Task Scheduling and Placement in Operating Systems for Dynamically Reconfigurable SoC", Proceeding of IFIP International Conference on Embedded and Ubiquitous Computing (EUC), LNCS 3824, pp. 489-498, 2005. [7] T.Marconi, Y. Lu, K.L.M. Bertels, and G. N. Gaydadjiev, "Online Hardware Task Scheduling and Placement Algorithm on Partially Reconfigurable Devices", Proceedings of International Workshop on Applied Reconfigurable Computing (ARC), pp. 306-311, London, UK, March 2008. [8] Y. Lu, T. Marconi, K.L.M. Bertels, and G. N. Gaydadjiev, "Online Task Scheduling for the FPGA-Based Partially Reconfigurable Systems", Proceedings of International Workshop on Applied Reconfigurable Computing (ARC), pp. 216-230, March 2009. [9] X. Zhou, Y. Wang, X. Huang, and C. Peng, "On-line Scheduling of Real-time Tasks for Reconfigurable Computing System", Proceedings of the International Conference on FieldProgrammable Technology (FPT), pp. 57-64, 2006. [10] X. Zhou,Y. Wang, X. Huang, and C. Peng, "Fast On-line Task Placement and Scheduling on Reconfigurable Devices", Proceeding of Field-Programmable Logic and Applications (FPL),

Bouraoui Ouni is currently an Associate Professor at the National Engineering School of Sousse. He has obtained his PhD entitled ‘Synthesis and temporal partitioning for reconfigurable systems’ in 2008 from the Faculty of Sciences at Monastir. He is obtained his university habilitation entitled ‘Optimisation algorithm for reconfigurable architectures’ in 2012. Hence, his researches interest cover: models, methods, tools, and architectures for reconfigurable computing; simulation, debugging, synthesis, verification, and test of reconfigurable systems; field programmable gate arrays and other reconfigurable technologies; algorithms implemented on reconfigurable hardware; hardware/software codesign and cosimulation with reconfigurable hardware; and high performance reconfigurable computing. Abdellatif Mtibaa is currently a Professor in Micro-Electronics and Hardware Design with the Electrical department at the National School of Engineering of Monastir and the Head of Circuits Systems Reconfigurable-ENIM-Group at Electronic and Microelectronic Laboratory. He received his Diploma in Electrical Engineering in 1985 and his PhD in Electrical Engineering in 2000. His current research interests include system on programmable chip, high level synthesis, rapid prototyping and reconfigurable architecture for real-time multimedia applications. He has authored/co-authored over 100 papers in international journals and conferences. He served on the technical programme committees for several international conferences. He also served as a co-organiser of several international conferences.



672


Classification of Brain Tumor Using Neural Network Bilal M. Zahran Abstract – Brain tumors classification in magnetic resonance imaging (MRI) is very important in medical diagnosis. Most of the current conventional diagnosis techniques are based on human experience in interpreting the MRI-scan for classification. This paper presents an automated method based on backpropagation neural network (BPNN) for classification of the MRI of a human brain. The proposed method utilizes wavelet transform (WT) as a feature extraction tool of the MRI. The proposed method follows two steps: feature extraction and classification. WT is first employed for decomposing the image into different levels of approximate and detailed coefficients and then these coefficients are fed into a BPNN for further classification and tumor detection. The proposed method has been applied on several MRI scans, and the results showed an acceptable accuracy of classification. Copyright © 2014 Praise Worthy Prize S.r.l. - All rights reserved.

Keywords: Brain Tumors Classification, Feature Extraction, Two-Dimensional Discrete Wavelet Transform, Back Propagation Neural Network

Neural computing provides an approach which is closer to human perception and recognition than traditional computing. Neural computing systems are adept at many pattern recognition tasks, so it is widely used in classification problems for image processing field. The discovery of the backpropagation algorithm leads to an increase of interest in neural networks. Feed forward multilayer networks trained with the backpropagation algorithm are still the most common kind today. The aim of our research is to recognize a tumor from a particular MRI scan of a brain image using digital image processing techniques, wavelet transform and neural networks. Several researches has been conducted in the field of classification of brain tumors using various automated techniques [4]-[10]. This paper is organized as follows; the second section will give an overview about neural network and wavelet transform. The third section will elaborate the proposed method. Section four will show the experiments and results done on classification process and will discuss the results. Finally section five will give a conclusion of the research.

Nomenclature MRI BP BPNN WT DWT 2D-DWT

Magnetic resonance imaging Backpropagation Backpropagation neural network Wavelet transform Discrete wavelet transform Two-dimensional discrete wavelet transform

I.

Introduction

A brain tumor is an intracranial mass produced by an uncontrolled growth of cells either normally found in the brain such as neurons, lymphatic tissue, glial cells, blood vessels, pituitary and pineal gland skull, or spread from cancers primarily located in other organs [1]. MRI has become a widely used method of high quality medical imaging, especially in brain imaging. MRI provides an unparalleled view inside the human body. It is widely used for detecting brain tumor. The field of image processing is a large one. At least it includes the following areas: image compression, image denoising, image enhancement, image recognition, feature detection and texture classification. In recent years a great effort of the research in field of medical imaging was focused on brain tumors segmentation and classification. Wavelet transforms have become increasingly important in image processing since wavelets allow both time and frequency analysis simultaneously [2]. WT decomposes the input image into a multiresolution space and feature vector for each pixel is computed. Neural networks are information processing systems which model the brain’s cognitive process by imitating some of its basic structures and operations [11].

II. II.1.

Neural Network and Wavelet Transform Artificial Backpropagation Neural Network

The application of neural networks to medical imaging has been increased recently. The purpose is to make use of the parallel distributed processing nature of neural networks to reduce computing time and enhance the classification accuracy. BPNN's are the most widely used neural network model as they can be applied to almost any problem that requires pattern mapping.



673

Bilal Zahran

In this research we used back propagation neural networks as a classifier. BPNN uses a supervised learning mechanism, and are constructed from simple computational units referred to as neurons [11]. Model of a neuron has three basic parts: input weights, a summer, and an activation function. The input weights scale values used as inputs to the neuron, the summer adds all the scaled values together, and the activation function produces the final output of the neuron. This description is illustrated in Fig. 1.

Step 1: Initialization Set all the weights and threshold levels of the network to random numbers uniformly distributed inside a small range. Step 2: Activation Activate the back-propagation neural network by applying inputs X1 (p), X2 (p),…., Xn (p) and desired outputs yd,1 (p), yd,2 (p),…., yd,n (p). And then calculate the actual outputs of the neurons in the hidden layer: ( )=

( )×

( )−

(2)

where n is the number of the inputs neuron j in the hidden layer and g is the activation function chosen. Step 3: Weight training Update the weights in the back-propagation network propagating backward the errors associated with output neurons. First, calculate the error gradiant for the neurons in the output layer:

Fig. 1. Model of an artificial neuron

where: aj : Activation value of unit j. wj,i: Weight on the link from unit j to unit i. ini : Weighted sum of inputs to unit i. ai : the output value. g : Activation function. Back propagation (BP) is one of the most famous training algorithms for multilayer neural networks

( )=

( ) × [1 −

( )] ×

( )

(3)

where: ek (p) = yd,k (p) – yk (p) Next calculate the weight corrections according to: ∆

( )=

×

( )×

( )

(4)

then update the weights at the output neurons according to: ( + 1) = ( )+∆ ( ) (5) Then repeat the process for the hidden layer as follows. First, calculate the error gradiant for the neurons in the hidden layer according to: Fig. 2. Architecture of backpropagation neural network

( )= BP is a gradient descent technique to minimize the error E for a particular training pattern. For adjusting the weight (Wi,j) from the ith input to the jth output, in the batched mode variant the descent is based on the gradient δE E ( ) for the total training set: δwij ∆

( )= − ×

+

×∆

( − 1)

( ) × 1−

( ) ×

( )×

( ) (6)

Next calculate the weight corrections according to: ∆

( )=

×

( )×

( )

(7)

Finally update the weights at the hidden neurons according to:

(1)

( + 1) =

The gradient gives the direction of error E. The parameters  and  are the learning rate and momentum respectively [12]. Here are the steps of the BP learning algorithm [12]:

( )+∆

( )

(8)

Step 4: Increase iteration p by one, go back to step 2 and repeat the process until the selected error criterion is satisfied or reaching the maximum number of iterations specified.



674

Bilal Zahran

II.2.

Wavelet Transforms

The wavelet transform is a tool that divides the signals into different frequency components, and then studies each component with a resolution matched to its scale. The continuous wavelet transform of the signal (t) is defined by: 

WT  ,a   a

1 / 2

*

  t   

 t     dt  a 

Ψ ( , ) =ψ (x)  ( y)

(11)

Ψ ( , ) =( y)ψ(x)

(12)

Ψ ( , ) =ψ (x)ψ( y)

(13)

where Ψ measures the horizontal variations (horizontal edges), Ψ corresponds to the vertical variations (vertical edges), and Ψ detects the variations along the diagonal directions. The 2D-DWT can be implemented using digital filters and down samplers. Fig. 3 shows the process of taking the onedimensional DWT of the rows of f (x, y) and the subsequent one-dimensional DWT of the resulting columns to produce three sets of detail coefficients including the horizontal, vertical, and diagonal details [3]. Fig. 4 shows the result of 2D-DWT decomposition of an image.

(9)

where a is a scale factor (also referred to as a dilation parameter) and  is a time delay. In wavelet analysis the basis functions  ,a  t  are oscillating functions and are called wavelets. The wavelets are scaled and translated versions of a prototype   t  , called the basic wavelet or mother wavelet or analyzing wavelet. The delay parameter  gives the position of the wavelet  ,a  t  while the scale factor a governs its frequency content. For a 1 the wavelet  ,a  t  is very much spread out and has mostly low frequencies. In wavelet analysis, we measure the similarity between the signal and the wavelet  ,a  t  for varying  and a which results in a set of wavelet coefficients which indicate how close the signal is to a particular basis function. The dilations by 1/ a result in several magnifications of the signal, with distinct resolutions [2].

Fig. 3. The analysis filter bank of the 2D-DWT

Approximation A2

II.2.1.

Multilevel 2-D Wavelet Decomposition

The discrete wavelet transform (DWT) uses multiresolution filter banks and special wavelet filters for the analysis and reconstruction of signals [2]. Images are analyzed and synthesized by 2D filter banks. The two-dimensional discrete wavelet transform (2DDWT) is used to yield good results in classification and segmentation of tumor from the brain MRI. In 2D-DWT, the image is represented by one approximation and three detail images, representing the low and high frequency contents image respectively. A two-dimensional scaling function,  (x, y), and three two-dimensional wavelet Ψ ( , ), Ψ ( , ) and Ψ ( , ) are critical elements for wavelet transforms in two dimensions [2]. These scaling function and directional wavelets are composed of the product of a one-dimensional scaling function  and corresponding wavelet ψ which are demonstrated as the following:

 (x, y) = (x)  (y)

Horizontal Detail H2

50

50

100

100

150

150

200

200

250

250 50

100

150

200

250

50

Vertical Detail V2

100

150

200

250

Diagonal Detail D2

50

50

100

100

150

150

200

200

250

250 50

100

150

200

250

50

100

150

200

250

Fig. 4. Result of wavelet decomposition of an image

Fig. 4 shows an example about 2D-DWT decomposition which results in three detail coefficients: intensity variations along columns (horizontal edges), intensity variations along rows (vertical edges) and intensity variations along diagonals.

(10)



675

Bilal Zahran

vector of feature values of the first abnormal image. The target output (T) vector will look like T=[1 1 1… -1 -1 -1 …] where 1 indicate normal image and -1 indicate abnormal image. To describe the performance of our proposed classifier, several terms that are commonly used along with the description of accuracy has to be cleared. They are true positive (TP), true negative (TN), false negative (FN), and false positive (FP).

III. Methodology Fig. 5 shows the steps of the proposed method.

Accuracy = (TN + TP)/(TN+TP+FN+FP)

(14)

= (Number of correct assessments)/Number of all assessments). The BPNN will be trained as explained in section II. According to the performance of the proposed method (Eq. (14)) the parameters of the WT (type of the wavelet and number of levels) and BPNN parameters (number of neurons in hidden layer, value of learning rate, activation functions for hidden and output layers) are tuned. After training process completes, the BPNN will act as a classifier for the MRI scan of the brain. Figures 6 show normal brain image and abnormal brain image (with tumor).

Fig. 5. Block diagram of the proposed method

Steps of the proposed method: Step 1: MRI preprocessing First, we read The MRI scan, next we resize the image (to obtain good classification, all images must have the same size), then, we convert the colored image to grayscale image where the image is a matrix of integers ranging from 0 to 255. These values specify shades of grey with 0 being pure black and 255 pure white. These integers can be specified using 8 bits (1 byte) for each pixel (matrix element).

(a)

Step 2: Feature extraction 2D-DWT is applied to the MRI scan which results in a decomposition vector C which consists of C = [ A(N) | H(N) | V(N) | D(N) | ... H(N-1) | V(N-1) | D(N-1) | ... | H(1) | V(1) | D(1) ]. where A, H, V, D, are row vectors such that: A = approximation coefficients. H = horizontal detail coefficients. V = vertical detail coefficients. D = diagonal detail coefficients. (b)

Step 3: Classification After repeating all preprocessing steps and finding feature extraction vectors for all images, BPNN is implemented as a classifier. The input vector (P) of the BPNN is a vector includes the feature values for all normal and abnormal images specified for training. P = [C1 C2 C3 ….Ca1 Ca2 Ca3 ….] where C1 is the vector of feature values of the first normal image and Ca1 is the

Figs. 6. (a) Abnormal brain MRI(with tumor); (b) Normal brain MRI

IV.


All running steps are conducted using Intel core2 due CPU T3200, 2.00GHz, 2GB RAM, and running on windows7 64-bit operating system. The code was written using MATLAB R2009b.



676

Bilal Zahran

TABLE IV RESULTS OF CLASSIFICATION WITH DB1 WAVELET AND 2 DECOMPOSITION LEVELS Normal Output Abnormal Output of image Of BP NN image BPNN No26 0.5236 ab15 0.7176 No27 0.8430 ab16 0.1683 No28 0.9105 ab17 -0.7257 No29 0.7392 ab18 0.0738 No30 0.4183 ab19 -1.0006 No31 0.7798 ab20 -0.9170 No32 0.6370 ab21 0.2600 No33 -0.8424 ab22 -0.2243 No34 1.0006 No35 0.9105

Our data consists of 35 normal images and 22 abnormal images (with brain tumor). The data set was divided into two parts: part one for training (25 normal, 16 abnormal) and part two for testing (10 normal, 8 abnormal). Both parameters of WT and BPNN were tuned to give the best results. After conducting several experiments, the best BPNN parameters, was: 5 neurons in hidden layer, sigmoid activation functions for hidden layer, Linear function for output layer and the learning rate was 0.5. Several experiments with various wavelets and decomposition levels are conducted and some of the results are shown in Tables I-IV. Results of Tables I-IV show the performance of the proposed algorithm. When the output of BPNN is greater than the threshold (0), then the image was classified as normal, otherwise it is classified as abnormal. The performance of the proposed method was measured according to Equation (14) to all experimental cases. The best accuracy had been achieved was 83.3%.

The algorithm classifies the normal images more efficiently than abnormal ones. The reason is that in some images the size of the tumor is too small and located in a small region in the abnormal image. In the best result the wavelet was db1 and 5 decomposition levels were used. An important factor which determines the performance of the algorithm is the number of decomposition levels used for WT.

TABLE I RESULTS OF CLASSIFICATION WITH HAAR WAVELET AND 5 DECOMPOSITION LEVELS Normal Output Abnormal Output image of BPNN image of BPNN No26 0.7676 ab15 0.9220 No27 -0.9847 ab16 0.3830 No28 0.8150 ab17 -0.4456 No29 0.9544 ab18 0.8908 No30 0.9458 ab19 -0.7993 No31 0.9733 ab20 0.9789 No32 0.9690 ab21 -0.9709 No33 -0.8911 ab22 -0.2099 No34 0.9947 No35 0.8150

V.

Conclusions and Future Works

In this paper, a backpropagation neural network was used as a classifier for brain tumors from MRI scans. First, after preprocessing the MRI, a two dimensional discrete wavelet transform was applied to the MRI, and then these features are fed to a BPNN to differentiate normal and abnormal images. The algorithm classifies the normal images more efficiently than abnormal ones. The reason is that in some images the size of the tumor is too small and located in a small region in the abnormal image. An important factor which determines the performance of the proposed method is the number of decomposition levels of the WT. The method has reached an acceptable accuracy (83.3%). As a future work, we suggest using an appropriate segmentation technique with our proposed algorithm and then using a suitable classifier. Also we can improve the results with using a proper noise estimation algorithm for MRI. As the higher accuracy of estimated noise pixels can improve the features of a filtered MR image [14].

TABLE II RESULTS OF CLASSIFICATION WITH DAUBECHIE 5 "DB5" WAVELET AND 5 DECOMPOSITION LEVELS Normal Output of Abnormal Output image BPNN image of BPNN No26 0.9342 ab15 0.6229 No27 0.9681 ab16 -0.9183 No28 0.9811 ab17 -0.0979 No29 0.9693 ab18 0.9680 No30 0.9582 ab19 1.0000 No31 0.9614 ab20 -0.9620 No32 0.9844 ab21 0.3175 No33 -1.0013 ab22 0.9376 No34 1.0012 No35 0.9811

References

TABLE III RESULTS OF CLASSIFICATION WITH DB1 WAVELET AND 5 DECOMPOSITION LEVELS Normal Output Abnormal Output image Of BPNN image Of BPNN No26 0.7676 ab15 -0.9220 No27 0.9847 ab16 0.2830 No28 0.8150 ab17 -0.4456 No29 0.9544 ab18 0.6908 No30 0.9458 ab19 -0.7993 No31 0.9733 ab20 -0.9789 No32 0.9690 ab21 -0.9709 No33 -0.8911 ab22 -0.3836 No34 0.9947 No35 0.9105

[1]

[2] [3]

[4]

[5]


W. Dou, , S. Ruan, , Y. Chen, D. Bloyet, and J.M. Constans, A framework of fuzzy information fusion for segmentation of brain tumor tissues on MR images, Image and Vision Computing, 2007. R.J.E. Merry, Wavelet Theory and Applications: A literature study, Eindhoven University of Technology. 2005 L. Prasad and S. S. Iyengar, Wavelet Analysis with Applications to Image Processing. Boca Raton, FL: CRC Press LLC, 1997, pp.101-115. T. Logeswari and M.Karnan, An Improved Implementation of Brain Tumor Detection Using Segmentation Based on Hierarchical Self Organizing Map, International Journal of Computer Theory and Engineering, 2010. C. Biradar, Shantkumari, Measurement based Human Brain


677

Bilal Zahran

[6]

[7]

[8]

[9]

[10]

[11] [12] [13] [14]

Tumor Recognition by Adapting Support Vector Machine, IOSR Journal of Engineering, 2013. L. Singh, R.B.Dubey, Z.A.Jaffery and Zaheeruddin, Segmentation and Characterization of Brain from MR Images Advances in Recent Technologies in Communication and Computing, International Conference, IEEE 2009. S. Roy, S. K. Bandyopadhyay, Detection and Quantification of Brain Tumor from MRI of Brain and its Symmetric Analysis, Department of Computer Science and Engineering, University of Calcutta, 2012. I. Maiti and M. Chakraborty, A New Method for Brain Tumor Segmentation Based on Watershed and Edge Detection Algorithms in HSV Colour Model, Computing and Communication Systems (NCCCS), National Conference, 2012. S. Ghanavati, J. Li, T. Liu, P. S. Babyn, W. Doda,G.Lampropoulos "Automatic brain tumor detection in Magnetic Resonance Images", Biomedical Imaging (ISBI), 9th IEEE International Symposium, 2012. M. Subashini and S. Kumar Sahoo ," Brain MR Image Segmentation for Tumor Detection using Artificial Neural Networks", School of Electrical Engineering, VIT University Vellore,2013. D.W. Patterson, Artificial Neural networks: Theory and Application, Prentice Hall.1996. M. Negnevitsky, Artificial Intelligence. Addison Wesley, 2nd edition. 2005. R. C. Gonzalez and R. E. Woods, Digital Image Processing 2/E. Upper Saddle River, NJ: Prentice-Hall, 2002, pp. 349-404. Sasirekha, N., Kashwan, K.R., A robust and hybrid Rician noise estimation scheme for magnetic resonance images, (2014) International Review on Computers and Software (IRECOS), 9 (2), pp. 247-254.

Authors’ information Al-Balqaa Applied University / Faculty of Engineering Technology, Department of Computer Engineering, P.O.Box 15008 Amman 11134 Jordan. Tel: 00962 785284732 E- mails: [email protected] [email protected] Bilal Zahran received the B.Sc degree in Electrical & Electronic Eng. from Middle East Technical University, Turkey, in 1996, the M.Sc degree in Communications Eng. from University of Jordan, Jordan, in 1999, and the PhD degree in Computer Information System (CIS) from Arab Academy for Banking and Financial Sciences, Jordan, in 2009. He is currently working as an Assistant Professor at department of Computer Engineering, Faculty of Engineering Technology, Al-Balqa’ Applied University, Jordan. His research interests include machine learning, data mining and optimization fields.



678


An Improved Semi Supervised Nonnegative Matrix Factorization Based Tumor Clustering with Efficient Infomax ICA Based Gene Selection S. Praba1, A. K. Santra2 Abstract – In cancer class discovery, tumor clustering becomes an active area of research. An exact identification of the type of tumors is essential for efficient treatment of cancer at the prior stage. Numerous techniques have been proposed and used to examine gene expression data, which clustering algorithms are inappropriate to many real-world problems whereas less amount of domain information only available. Adding the domain knowledge can help a clustering algorithm, to improve the quality of clustering result. In this paper, a semi-supervised non-negative matrix factorization (SS-NMF) structure is proposed for selected gene clustering and the genes are selected using proposed infomax ICA approach. Proposed system gives the tumor clustering result in terms of representing pairwise constraints on a small number of data objects that are to be clustered together while specifying whether “cannot” or “must”. Use of iterative procedure perform symmetric trifactorization data similarity matrix to conclude the cluster result. Finally, the experiments on the gene expression dataset verified that the proposed scheme can attain improved clustering results. Copyright © 2014 Praise Worthy Prize S.r.l. - All rights reserved.

Keywords: Clustering, Gene Expression Data, ICA, Semi-Supervised Non-Negative Matrix Factorization, Tumor

Representation of array for gene expression data becomes a popular technology for evaluation of thousands of genes expression simultaneously. These array based technology investigations have been demonstrated with possible usefulness of gene expression profiling for classifying tumors [4]. Rather than conventional morphological systems, gene expression profile gives extra information for the classification of tumors. Due to the fast development of microarray technologies, it can concurrently evaluate the gene expression level, makes the accurate, objective, and methodical analysis, and diagnoses of human cancers possible. A reliable and exact classification of the type of tumors is necessary for efficient treatment of cancer. Microarray data includes thousand of gene related information on every fragment and the number of data samples collected from gene is much lesser than the genes. So it is a characteristic “greater p, lesser n” problem [5], i.e., the number of predictor values p is larger than the number of samples n. It becomes difficult to use together analytical and interpretative points of view for standard statistical methods. For example, considering the samples with many variables may decrease the clustering accuracy and make the cluster rules complex to set. The inclusion of unrelated or noisy variables may too degrade the overall performances of the predictable cluster rules.

Nomenclature X and Y P(x) W ( ) ∈ × × ∈

Random Variables Probability Unmixing matrix Super-Gaussian distributions Cluster centroid Cluster indicator

I.

Introduction

The detection of cancer classes has usually been based on histomorphology. Discovery of cancer classes DNA microarrays have been efficiently used via clustering of the gene expression profiles. It has been shown that number of tumors can be clustered into similar groups based entirely on gene expression (mRNA) profiles [1]. It has been established that tumors that have equivalent histopathological appear might follow considerably different clinical courses with different responses to therapy and thus, incompetent diagnosis is often probable if the diagnosis is based mainly on morphological appearance. Microarray technology is experimental to classify tumor samples based on gene expressions and consequently has been generally used in system biology [2].Wide ranging of research works have been performed in earlier years while transforming the significance of tumor classification starting from morphological to molecular domain [3].



679

S. Praba, A. K. Santra

Regardless of these difficulties the clustering and classification methods from the areas of statistical machine learning have been applied to cancer identification using molecular gene expression data [6], [7]. In this paper, unsupervised clustering-based cancer class detection is proposed, as a substitute of supervised classification. Since the clustering approach doesn’t require priori the classes of training datasets, which are necessary by the supervised learning methods. Up-to-date several familiar unsupervised methods and extension of the methods such as Hierarchical clustering (HC), selforganizing maps (SOMs), nonnegative matrix factorization (NMF), and its extensions have been used effectively for cancer clustering [8]. In this paper, an Infomax ICA approach is proposed to select the genes and the selected genes are clustered by presented Semisupervised non-negative matrix factorization (SS-NMF).

II.

This novel hybrid approach integrates gene ranking, heuristic clustering analysis and wrapper technique to choose marker genes for tumor classification. In this approach, Li Jiangeng et al., initially the feature filter algorithm selects a group of top-ranked informative genes; then, in order to lessen the redundancy of the informative genes, a group of prototype genes are taken as the representative of the informative genes by heuristic Kmeans clustering; finally, SVM-FRE approach is utilized to choose a set of marker genes. ICA can reduce the effects of noise or artifacts on the signal and is efficient for separating mixed signals [18], [19]. mPCA and whether it should be called a multifaceted clustering algorithm or a dimensionality reduction algorithm. The stress-test the algorithm along two dimensions, the number of documents and their average size. This gave us the opportunity to evaluate mPCA methods more fully. This required scaling up the algorithm, although we have not yet desired to concern parallel processing (e.g., (Blei et al., 2002 [20])).Independent PCA models are set up for normal and cancer samples equally with the idea of Jolliffe’s PCA interpretation used to obtain feature genes for the microarray classification.

Literature Survey

A number of techniques have been presented in the literature to minimize the dimensionality of gene expression data [9]. A majority of the machine learning approaches have been used in cancer classification using microarray data [10]. Yendrapalli et al., [11] proposed a technique to estimate the impact of model selection on the performance of a number of SVM implementations to classify tumors. The issue of multiclass classification, particularly for techniques like SVMs, does not provide an easy solution. It is usually straightforward to build classifier theory and algorithms for two equally exclusive classes than for N mutually exclusive classes. Yendrapalli et al., utilized BSVM that constructs N-class SVMs [12]. A majority of the existing techniques for model selection utilize the leave-one-out (loo) related estimators which are regarded computationally costly. The author utilized Leave-one-out model selection for SVM (looms) that employs advance numerical techniques which results in efficient calculation of loo rates of different models [13]. Antai Wang et al [14] proposed a gene selection method for microarray data analysis using PCA with a modified gene number stop rule based on the strategy proposed by Asparoukhov, O et al. [15]. D. M. Witten et al. [16] introduced a Sparse PCA (SPCA) approach for sparse PCA data to estimate the principal components with sparse loading. The SPCA is built on the fact that PCA can be written as a regression optimization problem with a quadratic penalty, which is expensive to compute. Classification of microarray data is similar to fault diagnose of industrial process monitoring data with similar problem exists in the standard PCA. Li Jiangeng et al., [17] presented a new hybrid technique for choosing marker genes from gene expression data.

III. Methodology III.1. Infomax ICA for Gene Selection Recently, many successful ICA based applications of microarray data analysis methods are reported to extract gene expression modes [21], [22], [23], [24]. In this section, the details of the ICA model for gene expression data are discussed. An improved version of ICA called Infomax ICA approach. In this work, Infomax ICA is used to select the genes and subset of tumor genes. An Infomax ICA method uses a mutual information gain to select the genes. Infomax ICA Mutual information is definite for a pair of two random variables X and Y is defined as below: ( ; )=

( )− ( | )

where ( | ) the conditional entropy of X taken Y on certain is values y and is the conditional entropy of X. Conditional entropy of x is defined as below: ( | )=

( , )− ( )

where ( , ) is the joint entropy of and and ( ) is the entropy of Y. Formally, entropy for a given random variable x and y is defined as: ( )=−


( )

( )


680


( )=−

( )

( , )=−

expression profiles by a new set of basis vectors. This result comes from following assumptions. Initially the gene expression profiles are evaluated by a combination of hidden variables, which are named as “expression gene of modes “(eigengenes). Following, the genes’ responses to those variables can be estimated by linear functions. Gene expression profile q of Y matrix is defined by all of the eigengenes in the rows of X and by its linear independent influences on the expression profile q th row of W eigengenes. Finding the set of good basis profiles to characterize gene expression data thus the subsets of genes relevant to cell samples classification can be selected. Also the main objective of this work is to choose the subset of genes that could be relevant for cell clustering. The selection of gene is obtained by Infomax ICA. The proposed gene selection process is based on a ranking of the p genes. This ranking process is introduced as follows:

( )

( , )

( , )

,

where P(x) denotes the probability of X in the state x i,e gene. Entropy is used to measures the value of uncertainty. The lower value of information gives detail result of system. Consequently, going back to, ( ; ) mutual information can be shows that decrease of uncertainty about variable X following the observation of Y. By representation of Infomax that seeks to reduce mutual information and correspondingly searching for components that are maximally independent. After the upadatation of ( ; ) result proposed the following algorithm to compute the unmixing matrix W (called InfoMax):

1. Initialize random variables of genes which is denoted as W. ( + 1) = 2. ( ) + ( ) + ( )( − ( ) ) ( ) 3. If not converged, go back to step 2 and choose another genes for the further procedure

1. Initialize (0) (e.g. random) ( + 1) = 2. ( ) + ( ) + ( )( − ( ) ) ( ) 3. If not converged, go back to step 2. where stands for a given approximation step, ( ) is defined as the size of the steps for the unmixing matrix updates ie usually exponential function or a constant value. ( ) a nonlinear function normally selection according to the type of gaussian distribution(super or sub), the identity matrix of dimensions × and the transpose operator. In the case of super-Gaussian distributions, ( ) is normally denoted as: ( )=

By this way, the genes and the subset of the gene are selected for further processing. The selected genes are given to the SS-NMF approach for clustering purpose in class discovery. III.2. SS-NMF for Gene Clustering Dimensionality reduction of gene expression data is performed by using NMF. NMF can be a well-organized method to recognize different molecular patterns and an authoritative tool for class discovery. Brunet et al. [6] established the capability of NMF to restore the efficient biological information that is related to the microarray data for cancer classification or clustering. But normal NMF cannot control the scattered of the decomposition, and thus, it cannot always yield a parts-based demonstration. In this paper, a Semi-supervised nonnegative matrix factorization (SS-NMF) approach is proposed to cluster the selected gene expression data from Infomax. Given a non-negative matrix of size × , where each column of non negative matrix corresponds to a selected genes in the -dimensional space result from infomaxICA and a positive integer < { , }, If find two non-negative matrices W and H where ∈ ℝ × and ∈ ℝ × , so that the non negative matrix A is followed by ≈ . NMF problem can be solved by using the optimization function is defined as follows:

ℎ( )

and for sub-Gaussian distributions, ( ) is denoted as: ( )=

−

ℎ( )

The package InfoMax.nb is an implementation of this algorithm. Gene Selection using Infomax ICA Gene expression dataset can be represented by a p × n matrix ( ≫ ), whose element is the expression level of the gene in the as say (1 ≤ ≤ , 1 ≤ ≤ ). Two dimensional vectors are used to represent the gene samples in the dataset there are n dimensional vector and p dimensional vector. The dimensional vector , i.e., the row of , denotes the expression profile of the gene. On the other hand, the pdimensional vector , i.e., the column of X, is the snapshot of the as say (cell sample). Consider that the data samples have been preprocessed with p gene samples with n samples and normalized, i.e., each cell sample has zero mean and unit standard deviation. In the Infomax ICA model for gene expression data in ICA model is defined as Y=WX represents the gene

,


( , )≡

1 ‖ − 2

‖ , . .

,

≥0


681


where ∈ ℝ × is a basis matrix, ∈ ℝ × is a coefficient matrix, ∥. ∥ is the Frobenius norm and , ≥ 0 means that all elements of and are nonnegative. Dimensionality reduction of gene expression data is done by using integer k Tmin then (RTT < RTTth) && (SCPU> Sth) && (C(t) > Cth) elseif tresxy (Ji) < Tmin then (RTT < RTTth) || (SCPU> Sth) || (C(t) > Cth) end if endif if Type(Ji) = Jcomm if tresxy (Ji) > Tmin then (Mo_P < Tth) && (RSS>Rth) elseif tresxy (Ji) < Tmin then (Mo_P < Tth) || (RSS>Rth) end if endif } …. ….

III.4.2. Communication-Focused Jobs (Jcomm) These are characterized by the large quantity of data communicating among processors. These data appears to be frequent large messages. The application includes telemedicine, disaster management and scientific collaboration [14]. The steps involved in the job scheduling are as follows. The client sends a resource requisition to the Grid Controller. The Grid Controller forwards the requested message to the proxy server. The proxy server fetches the (R_req) and analyzes the job category. It estimates the response time of all the nodes in the client-set for the job as per equation (5). It then sorts the list of the suitable resources for the job according to the response time in ascending order. The resource allocation for the above mentioned jobs depends on the parameters such as round trip time (RTT) and CPU speed (SCPU) and capacity(C(t)), mobility (Mo_P), received signal strength(RSS) and reliability(R) (Estimation are briefed in sections (3.2.1-3.2.6). If the job is computation focused and response time is larger than a minimum threshold, then the proxy monitors the node within each client-set for its round trip time, CPU speed and capacity. Since the response time is larger, the proxy selects a node that has shorter Round Trip Time(RTT < RTTth), greater CPU speed (SCPU> Sth) and high CPU (C(t) > Cth) capability. If the response time is lesser than the minimum threshold, the job can be assigned to a node which has either shorter Round Trip Time. (RTT < RTTth) or high CPU speed(SCPU> Sth) or high CPU (C(t) > Cth) capability. Where RTTth, Sth, Cth are threshold value for Round Trip Time, CPU speed and CPU capacity respectively. The threshold values are based on Information Providers. If the job is communication focused, response time is larger than a minimum threshold then the proxy monitors the node within each client-set for its RSS and mobility. It then assigns the job to a node with lesser mobility (Mo_P < Th) and with a greater signal strength (RSS>Rth). If the response time is lesser than the threshold, the job can be assigned to node which has either lesser mobility (Mo_P < Th) or greater signal strength (RSS>Rth). Upon selecting the node, the proxy server sends the reply message (R_rep) about the suitable node

IV. IV.1.

Simulation Results

Simulation Model and Parameters

In this section, we examine the performance of our Mobility Aware Load Balanced Scheduling Algorithm (MALBS) with an extensive simulation study based upon the Network Simulator Version-2 (NS-2) [18]. The simulation topology is given in Fig. 2. We compare our results with balanced scheduling algorithm BSA technique [6]. Various simulation parameters are given in Table I.

Fig. 2. Simulation Setup



690

S. Vimala, T. Sasikala

TABLE I SIMULATION SETTINGS 9 6 3 1000 X 1000 802.11 250m DSDV 50 sec CBR 512 250,500,750 and1000Kb 5m/s to 25 m/s 0.660 w 0.395 w 0.335 w 10.1 J

3.5 3 time (sec)

Mobile Nodes Users Clusters Area Size Mac Radio Range Routing Protocol Simulation Time Traffic Source Packet Size Maximum Rate Speed of mobile nodes Transmit Power Receiving Power Idle Power Initial Energy

Load Vs Response Time

2.5 2

BSA

1.5

MALBS

1 0.5 0 250

500

750

1000

Load (kb)

Fig. 4. Load Rate Vs Delay

Load Vs Packet Drop 600

Performance Metrics

Pkts

IV.2.

In our experiments, we measure the following metrics.  Delay: It measures the average end-to-end delay occurred while executing a given task.  Received Bandwidth: It is the ratio of the number of bits successfully transmitted to the receiver per sec.  Drop: It is the total number of packets dropped during the data transmission. IV.3. IV.3.1.

BSA

200

MALBS

0 250

500

750

1000

Load (Kb)

Fig. 5. Load Vs Drop

It considers CPU speed, RTT and capacity of nodes for computation intensive jobs.

Results

Based on Load IV.3.2.

In the initial experiment, we vary the maximum load of the job requests, by varying the rate as 250,500,750 and 1000Kb. Fig. 3 shows the throughput received in case of both the techniques when the load is increased. From the figure, we can see that the received throughput of MALBS is 20% higher than the existing BSA technique. Fig. 4 shows the response time measured in case of both the techniques when the load is increased. From the figure, we can see that the response time of MALBS is 54% less than the existing BSA technique. Fig. 5 shows the packet drop measured in case of both the techniques when the load is increased. From the figure, it can be seen that the packet drop of MALBS is 60% less than the existing BSA technique.

Based on Speed of Mobile Nodes

In the initial experiment, we vary the speed of the mobile nodes, by 5 to 25 m/s and measure the performance metrics. Node Speed Vs Throughput

Throughput (Mb/s)

0.25 0.2 0.15

BSA MALBS

0.1 0.05 0 5

10

15

20

25

Speed(m/s)

Fig. 6. Node Speed Vs Received Throughput

Load Vs Received Throughput

Node Speed Vs Response tim e

0.3 0.2

2

BSA Time (sec)

Throughput (Mb/s)

400

MALBS

0.1 0 250

500

750

1000

1.5

BSA

1

MALBS

0.5 0 5

Load (Kb)

10

15

20

25

Speed (m /s)

Fig. 3. Load Vs Received Throughput

Fig. 7. Node Speed Vs Delay



691


[4]

Fig. 6 shows the throughput received in case of both the techniques when the node speed is increased. From the figure, we can see that the received throughput of MALBS is 30% higher than the existing BSA technique. Fig. 7 shows the response time measured in case of both the techniques when the node speed is increased. From the figure, we can see that the response time of MALBS is 82% less than the existing BSA technique. Fig. 8 shows the packet drop measured in case of both the techniques when the node speed is increased. From the figure, it can be seen that the packet drop of our proposed MALBS is 41% less than the existing BSA technique. It considers both RSS and mobility of the nodes for communication intensive jobs.

[5]

[6]

[7]

[8]

Node speed Vs Packet Drop 70

[9]

60

Packets

50 BSA

40

[10]

MALBS

30 20 10

[11]

0 5

10

15

20

25

Speed(m/s)

[12]

Fig. 8. Node Speed Vs Drop

V.

Conclusion

[13]

In this paper, we have proposed a mobility aware load balanced scheduling algorithm for mobile grid environment. In this technique, two job categories such as computing-focused and communication-focused jobs are taken into consideration. The server allocates the resources with shorter round-trip-time (RTT) and high CPU speed and capacity to the computing-intensive jobs. The communication-intensive jobs are allocated to resources with low mobility and high reliability. The main advantage of this approach relays on job scheduling based on nodes execution capability. This approach provides an exceptional balanced job scheduling across the entire client set in mobile grid network. By simulation results, we have shown that the proposed approach minimizes the network overhead and increases the network performance.

[14]

[15]

[16]

[17]

[18] [19]

[20]

References [1]

[2]

[3]

Li Chunlin and Li Layuan, “Energy constrained resource allocation optimization for mobile grids”, Elsevier, Journal of Parallel Distrib. Comput., 70, pp 245-258, 2010. Joanna Kołodziej, Samee U. Khan, Lizhe Wang, Dan Chen and Albert Y. Zomaya, “Energy and Security Awareness in Evolutionary-driven Grid Scheduling”, Springer, Evolutionary Based Solutions for Green Computing Studies in Computational Intelligence, vol 432, pp 95-138, 2013. S. Stephen Vaithiya and S. Mary Saira Bhanu, “ZONE BASED JOB SCHEDULING IN MOBILE GRID ENVIRONMENT”, International Journal of Grid Computing & Applications (IJGCA) vol.3, No.2, June 2012.

[21]

[22]


Preetam Ghosh, Nirmalya Roy and Sajal K Das, “Mobility-Aware Efficient Job Scheduling in Mobile Grids”, IEEE Computer Society, Seventh IEEE International Symposium on Cluster Computing and the Grid, 2007. Ashish Chandak, Bibhudatta Sahoo and Ashok Kumar Turuk, “An Overview of Task Scheduling and Performance Metrics in Grid Computing”, IJCA Special Issue on 2nd National ConferenceComputing, Communication and Sensor Network (CCSN) (1):, pp 30- 33, 2011. JongHyuk Lee, SungJin Song, JoonMin Gil, KwangSik Chung, Taeweon Suh1, and HeonChang Yu. “Balanced Scheduling Algorithm Considering Availability in Mobile Grid”, Advances in Grid and Pervasive Computing (2009):, pp 211-222, 2009. Ghosh, Preetam, Nirmalya Roy, and Sajal K. Das. "Mobilitybased Cost-effective Job Scheduling in an IEEE 802.11 Mobile Grid Architecture." The University of Texas at Arlington, 2007. Chin, SungHo, Taeweon Suh, and HeonChang Yu. "Genetic Algorithm based Scheduling Method for Efficiency and Reliability in Mobile Grid", proceedings of the 4th International Conference on Ubiquitous Information Technologies & Applications, ICUT'09., IEEE, 2009. Cilku, Bekim, and Aksenti Grnarov. "New algorithms for efficient scheduling in Grid Ad-Hoc networks", proceedings of the ITI 2009 31st International Conference on Information Technology Interfaces, Jun, pp 22-25, 2009. Lin, Chih-Kuang, Vladimir Zadorozhny, and Prashant Krishnamurthy. "Grid-based access scheduling for mobile data intensive sensor networks", 9th International Conference on Mobile Data Management, MDM'08., 2008. Katsaros, Konstantinos, and George C. Polyzos., "Evaluation of scheduling policies in a mobile grid architecture", . International Symposium on Performance Evaluation of Computer and Telecommunication Systems, SPECTS 2008, IEEE 2008. Noureddine Kettaf, Hafid Abouaissa, Thang Vuduong† and Pascal Lorenz, “A Cross layer Admission Control On-demand Routing Protocol for QoS Applications”, IJCSNS International Journal of Computer Science and Network Security, vol 6, No.9B, Sep 2006. D. Adami, C. Callegari, S. Giordano, M. Pagano, “A Hybrid Multidimensional Algorithm for Network-aware Resource Scheduling in Clouds and Grids”, ICC, pp 1297-1301, IEEE 2012. Javier Bustos, Denis Caromel, Mario Leyton, and José M. Piquer, “Load Information Sharing Policies in Communication-Intensive Parallel Applications”, in ISSADS, Lecture notes in computer science, Springer, 2006. Said Fathy El-Zoghdy, “A Hierarchical Load Balancing Policy for Grid Computing Environment”, I. J. Computer Network and Information Security, vol 5, pp 1-12, 2012. T. Altameem, “On the Design of Job Scheduling Strategy Using Agent Replication for Computational Grids”, International Journal of Computer Science and Network Security (IJCSNS), vol.11, No.3, March 2011. A.R. Sandeep, Y. Shreyas, Shivam Seth, Rajat Agarwal, and G. Sadashivappa “Wireless Network Visualization and Indoor Empirical Propagation Model for a Campus WI-FI Network”, World Academy of Science, Engineering and Technology,18 2008. Network Simulator: http:///www.isi.edu/nsnam/ns Kong, F., Hao, H., Zuo, J., Classification and dynamic fuzzy clustering of mold resources based on mold manufacturing grid platform, (2012) International Review on Computers and Software (IRECOS), 7 (5), pp. 2447-2452. Li, X., Hu, Z., Yan, C., A queuing time aware dynamic Grid workflow scheduling model based on repairable queuing system, (2012) International Review on Computers and Software (IRECOS), 7 (5), pp. 2608-2615. Prabhakar Telagarapu, L. Govinda Rao, D. Srinivasa Rao, P. Devi Pradeep, Analysis of Mobile User Identification Inside the Buildings, (2011) International Journal on Communications Antenna and Propagation (IRECAP), 1 (2), pp. 196-203. A. El Fallahi, Implementation of the UMTS Technology in the GSM Existing Network: Capacity/Interference Optimization, (2012) International Journal on Communications Antenna and Propagation (IRECAP), 2 (6), pp. 372-376.


692


Authors’ information S. Vimala she received A.M.I.E degree in Computer Science and Engineering in 2001 from The Institution of Engineers (India) in 2001 and M.Tech degree in Computer Science and Engineering with First Class with Honors from Dr.MGR University., Chennai in 2008. She is presently doing her research in the area of Grid Computing under Anna University, Chennai., India. She has 13 years of teaching experience in various Engineering colleges. She was also awarded “Suman Sharma Award” by The Institution of Engineers (India). Her research interests include computational grid, distributed computing and mobile grid. She is a member of the IEEE. Dr. T. Sasikala she received B.E. degree in Computer Science and Engineering with First Class in 1991, M.E. degree in Computer Science and Engineering with 7thrank from Madras University., India in 2001 and Doctor of Philosophy in Sathyabama University in the year 2009 respectively. She has 21 years of teaching experience in various Engineering Colleges. She is currently working as Principal in SRR Engineering College, Chennai., India. She has published over 30 papers in various journals and Conferences. She was also awarded as “The best teacher” by Sathyabama University., India. Her research interests include Wireless networks, Grid Computing, high speed computer networks, Network security and Data mining.



693


A Novel Localization Approach for QoS Guaranteed Mobility-Based Communication in Wireless Sensor Networks P. Jesu Jayarin, J. Visumathi, S. Madhurikkha Abstract – In wireless sensor networks the Nodes are mobility enabled, since the communication among nodes fully depend on the distance between nodes and time taken to transmit the data packets. Any routing protocols suggesting the shortest path based route formation is the best concept for fast and safe transmission and it improves the QoS of the WSN in terms of speed and accuracy. Without locating the nodes and the accurate distance, the shortest path finding is inefficient. In this paper a novel Localization approach is introduced to localize the nodes accurately with the help of Beacon Nodes in the network. The simulation results produced better performance than the existing systems. Copyright © 2014 Praise Worthy Prize S.r.l. - All rights reserved.

Keywords: Location Estimation, Localization, Mobility, Quality of Service, Shortest Path Communication, Wireless Sensor Network

Non-conformity of cell traffic to the prior knowledge evidently degrades the performance. Hence with the change of network traffic, application of dynamic scheme, adapting the channel value is better. The time of channel reservation for the incoming handoff calls needs to be addressed in order to determine either an optimal or near optimal reservation value channel at any specific time. The sufficient reservation is made at a time that can be utilized in the near future; which will certainly achieve an enhanced performance [3]. If handoff calls do not utilize reservation, resources will be wasted unnecessarily in blocking new calls. User mobility during the call connection causes handoff calls. Hence, there is a need for better reservation scheme depending on the mobility pattern of the user. Several factors such as destination of the user, cellular network layout, network’s current traffic condition are largely responsible for determining such mobility patterns. Detailed characterization of the pattern of mobility for every individual user is difficult and is not useful either, as call performance is a resultant collective outcome of total users in the network. Instead, using the statistical property of the mobility [4] of the users would be more useful. Based on their velocities users can be classified into two groups: high speed (vehicular users) and low speed (pedestrians) users. On an average high speed user’s cell dwell time is shorter than that of low speed user. These classifications help us to predict the probability handoff [5] of each group and accordingly make reservations. The propagation of wireless sensor network communication tools has been permitted in the progress of WSN, which comprises a huge number of minor and inexpensive sensors with limited resources, such as calculating, announcement, stowage and energy [8].

Nomenclature QOS WSN TDOA AOA MS SINR RSS GPS Ri Prob Acc LOC_REQ LOC_RES RES REQ Ref E Dacc Z

Quality of Service Wireless Sensor Network Time Difference of Arrival Angle of Arrival Mobile Station Signal-to-Noise pulse Ratio Received Signal Strength Global Positioning System Reference Index Probability Accuracy Location Request Location Response Response Request Reference List Error Distance-Accuracy Reference Node

I.

Introduction

A cellular communication system [1], [2] is considered where we have a provision for N channels in each cell. A number of channels [2] among the existing N channels can be reserved for handling the incoming handoff calls in order to prioritize the handoff calls. Prior knowledge of the traffic cell [1]-[7] and requirement of call blocking are used to determine the number of channels in the conventional cut of priority scheme or guard channel scheme. Channels do not change during the operation of the system. Manuscript received and revised March 2014, accepted April 2014


694

P. Jesu Jayarin, J. Visumathi, S. Madhurikkha

These sensor nodes are talented to sense, measure and gather uncooked data from the background, completes simple computations and then transmit only the requisite and partially administered data to the node answerable for combination [9]. Wireless sensor network ([34]-[36]) is deployed extensively in areas such as surveillance, health monitoring [11], human action detection, military [10] and disaster management [12] and hazardous environments [13]. This kind of applications requires that the position of the nodes must be determined. In most of the scenarios node position information gives a critical situation, such as data centric storage based applications [14]. Moreover, the algorithm fulfills several secondary design objectives: self-organizing nature, simplicity, robustness, localized processing and security.

II.

requires special sensor nodes termed as beacons [17]. In another way that without the use of beacon nodes [18] the location of each nodes are estimated are with the help of nodes related due to some geometric features. There is an algorithm called incremental algorithm [19] which starts with a less number of beacon nodes. Once their location is estimated their points are taken as reference points and they increase the x, y value and find the other nodes location. Another approach, called concurrent approach [20], [21] takes pair of nodes and by their communication the nodes to share their measurements and find the location. Fine-grained method [22] uses accurate information in the position estimation which measure the distance by using beacons RSS or the ToA – [time of arrival] procedure. Coarse-grained approaches [23] utilize less information with the help of rough techniques like hop count and measure the distance to beacons. A centralized approach [24] needs global knowledge to measure all the existing data with the data provided to the nodes in the distributed network. In example [25] an algorithm is proposed and it needs a triangle shape based beacon placed in a particular location, It assumes if the localization algorithm is independent of global infrastructure and beacon placement [14], else it would increase the computation cost which is clearly stated in [26]. Some of the localization algorithms [28], [29] contain two important main steps: initialization step, where a node get a rough estimation of its location and the refinement step, where each node repeatedly broadcasts its initial position, through which the refined current position is estimated. There are several algorithms introduced to overcome the problems in the existing approaches which are given in [30]. A portion of the sight and sound sensors consolidated with scalar sensors [31] to enhance the QOS as far as ongoing interactive media requisitions. These sensors are utilized to make another era of MAC conventions for WMSN focused around Ieee802.11e. Improved Distributed Channel Access convention [32] utilized as a part of MAC layer for enhancing QOS in WSN as far as VOIP based correspondence. In [33] the creator rolled out an improvement in system topology to keep up a concentrated MAC for MANET and attempted to evade self-centered hubs in the system. Various algorithms and techniques are analyzed to motivate and to compare the proposed approach’s efficiency. It means that the problems in the existing papers and the methods should be rectified and compared for better performance.

Related Work

For locating the users have to merge both TDOA [Time Difference of Arrival] + AOA [Angle of Arrival] [7] have to be merged. The home BS [Base Station] for an MS [Mobile Station], renders services to the target MS (to be located). All the adjacent BSs, which contain signals having the signal-to-interference-plus noise ratio (SINR) above a specified threshold at the MS, can be involved in the process of an MS location. The MS constantly monitors the levels of such forward pilot channel signals captured from the neighboring BSs and reports the information about the signals that cross the predefined set of thresholds to the network. The MS receiver through the cross correlates can measure the TDOA [7] between the signals either from home BS or other BSs. Adaptive antenna arrays in third generation cellular communication systems, for the purpose of radio transmission are proposed, which facilitate the tasks of initial acquisition, tracking of time, coherent reference recovery of Rake-receiver, and measurements of power-control for the MS. Availability of an adaptive antenna array [7] at the BS site enables the home BS to dedicate a spot beam to an individual MS under its scope by dynamically readjusting the antenna direction pattern in accordance with the movement of MS in such a way, as to furnish the signal from the MS arriving based on azimuth angle. Enhanced location accuracy can be achieved by using the AOA [7] measurement in conjunction with TDOA measurement. The prime assumption here is that at any given point of time, forward-link pilot signals can be received by the MS to be located both from its home BS and at least from one adjacent BS. Localization techniques entirely differ from one another in terms of various features [15], like the way of gathering input data, the node state in the sensor nodes and the application node answerable for location estimation. In [16] the different types of algorithms are categorized. Using a GPS receiver or manual configuration of nodes the location is discovered, and it

III. System Model The localization method needs the information from node to be shared with the other node. By deploying the information of one node into the other node, the accuracy of the location estimation, reduction of the communication information, and requirement of the computational services are shown in Fig. 1.



695

P. Jesu Jayarin, J. Visumathi, S. Madhurikkha

Level-3:Sophisticated position estimation o For [j=1 to C(Si))  − − ,| , = |  ( , > )then [further == true); break. o [ ℎ = ]  ( =1 ( ))  − − ,| , = |  > ) Then [eliminate rj]. ,  Estimate the refined position as shown in 2 o  =

Beacon Node Normal Node

Level-4: o = ∑∈ o ( <  ̂ ( 
topic(+m;d)

Fig. 2. Steps of PROCEOL Technique

(5)

In simple MLN model have the formulas of linking word to page classes and page classes to the classes of linked pages. The word-class rules were described as:

IV.3. Co-Occurrence Analysis Co-occurrence analysis is broadly used in different forms of research concerning the domains of ontologies, text mining, data extraction, knowledge extraction and content analysis etc. Commonly its aim is to find similarities in sense between word pairs and similarities in meaning within word patterns, also in order to discover latent structure of mental and social representations. Co-occurrence data is said to be the D which have the terms that are similar for particular context. Co-occurrence analysis challenges to identify lexical units that tend to occur together for purposes ranging from extracting related terms to discovering implicit relations between concepts.

Has(p,g) => Class(p,m)

(6)

-Has(p,g) => Class(p,m)

(7)

Here pair of (word, class) is described as p is represent as page, g represents as word, and m represents as class. Linked page classes were related by the formula of: Class(p1,m1) ʌ LinksTo(p1, p2) => Class(p2,m2)

(8)

Therefore, system needs a rule for each (word, class) pair and a rule stating that, given no evidence, a document does not belong to a class: Topic(m; d). Evidences are collected from the true pairs of the predicates which are the Concept (concepts). Here the concept is described as the Concept (ck) The performance of inference process is to gather the truth values of the possible trainings of the R(concept; word) predicate which means that a word is a lexical realization of a concept, i.e. (wj ; ck) Є R, based on the evidence. Here the document is represented as d, and the word or term represented as g. The Term g is collected as {g1,g2,g3,g4,….. , gN}, N be the number of corpus. The evidence is composed of a set of trainings of the HasWord(document; word) predicate, which means that a term is present in a document and the Depends(word; word; dependency) predicate, which means that a word governs another word through the specified syntactic dependency, which is described as:

IV.4. Concept Identification Concept Identification process is performed by the technique of Markov Logic Network. In order to implement the concept identification using Markov Logic Network, system presents the process of Weight Learning and the Inference. Learning of weight contains the three methods, which are Discriminative Learning, Generative Learning and Imitative Learning. Generative learning is at the main process of many approaches to pattern analysis and recognitions, artificial intelligence, and perception and provides a rich framework for imposing structure and prior knowledge on a given problem. Discriminative Learning is the maximizing of conditional log likelihoods. Imitative Learning is process that presented as another variation on generative modeling which also learns from exemplars from an observed data source. Here system uses the learning weight of discriminative learning, because the discriminative prediction task outperforms the usual generative approach which maximizes the joint likelihoods of all predicates. For make the discriminative learning method Voted Perceptron, Conjugate Gradient, and Newton’s Method are used for it. For all the learning weight finally provides the inputs into the Expectation Maximization which is an iterative method used to find maximum likelihood estimates of parameters in probabilistic models.

depends(g3,g1, dep) ^ depends(g3,g2,dep) ^ R(c,g1) => R(c,g2)

(9)

To make the complex relations in Markov Logic Network inference are used us probabilistically. There are two basic types of inference: maximum a posteriori (MAP)/most probable explanation (MPE) inference that finds the most probable state of the world consistent with some evidence, as well as conditional/marginal probability inference that finds the conditional/marginal distribution of a formula or a predicate.



721

K. Karthikeyan, V. Karthikeyani

In this PROCEOL technique, system uses the MCSAT algorithm [25], it is the process combines the MCMC and SampleSAT. SampleSAT combines the MaxWalkSAT and the simulated annealing. WalkSAT is repeatedly flipping a variable in a random unsatisfied clause. It is described by with probability p and with probability 1-p. Like that MaxWalkSAT is the highly non-uniform sampling that is extended to the weighted satisfiability problem. Simulated annealing is the technique of Metropolis algorithm, which is also been used in algorithms for approximate counting of large data sets. This process is mostly applied to optimization problems. A Markov Chain Monte Carlo process is used for computing marginal and conditional probabilities in Markov Logic Networks. This technique is performed by the combination of Markov Chain Monte Carlo with the MC-SAT algorithm. MC-SAT selects a random gratifying assignment of a random subset of the currently satisfied phrases. This allows MC-SAT to handle the mix of hard and soft constraints which are often present in MLNs. Here Y(0) describes the hard constraints. k is the samples that we are taken for the process. R is the random subset of clauses. It can be shown that the set of samples from MC-SAT combines to the correct distribution as long as the satisfying assignments are selected uniformly at random. MC-SAT provides the probabilistic inference. Then completion of this process it performs the process of Probabilistic Latent Semantic Analysis (PLSA).

corpus with probability p(d). And to analyse the latent class in the document, the class is said to be the z is described as z Є Z = {z1,z2,z3,z4,… ,zK}, Let K has to be specified a priori with probability p(z/d) and finally the probability for the word, to generate a word g is described as g Є G = { g1,g2,g3,g4,… ,gM}, Let M be the number of distinct words from the corpus with probability p(g/z), where:

Pg / d  

 Pg / z Pz / d 

(10)

z

Fig. 3 explains the process of the contents that are described above. The joint probability between a word and document, p(d,g) is presented by

P  d ,g   P  d  P  g / d   P d 

 Pg / z Pz / d 

(11)

z

and using Bayes’ rule can be written as:

P  d ,g    P  g / z  P  d / z  P  z 

(12)

The constraints of the model are p(g/z) and p(z/d); its number is (M-1)K, respectively N(K-1), which means that the total number of constraints grows linearly with the size of the corpus and the model becomes prone to over fitting. The constraints (i.e. parameters) can be calculated from the likelihood maximization, by identifying those values that maximize the predictive probability for the observed word occurrences. The predictive probability of PLSA mixture model is represented by p(g/d), so the objective function i.e. likelihood function is presented by: L n  d ,g  log P  d ,g  (13)

MC-SAT algorithm: Y(0) ← Hard clauses random solution for k ← 1 to total number of samples that are taken in the processing (whole data sets) R ← ø for probability of 1-exp(-wi) add Ci to R sample Y(k) uniformly random solution satisfying R endfor

 d

g

Thus this process is used for the concept extraction calculations. The PLSA method calculates the relevant probability distributions by selecting the model parameter values that maximize the probability of the observed data, observed data is said to be the likelihood function. Expectation Maximization (EM) algorithm is a standard algorithm used for maximum likelihood estimation. Thus the final results of the concept extraction are Labelled TS, R and S, i.e. term extraction with concept extracted data.

IV.5. PLSA PLSA (Probabilistic Latent Semantic Analysis) goal is to identifying and discriminating between different contexts of word usage without recourse to a dictionary or thesaurus. This model is implemented by the following process, that are:  Document Selection,  Probability for Latent class,  Probability for the word.

V.

Experiments

In this section, we conduct an extensive set of experiments to examine the concept extraction performance of the proposed PROCEOL framework by comparing with a state-of-the-art technique. The

Document Selection Process consider as a document d is described as d Є D = {d1,d2,d3,d4,………. ,dN}, Let N be their number, defined by the size of our given



722


proposed PROCEOL consists of the process of noise removal with the concept extraction. That provides the accurate extraction results. This extraction results and other dataset results are shown in below. V.1.

Extraction. Concept Extraction process is differs from the PRECE as compared with the PROCEOL. Some additional features are added in our proposed PROCEOL method because of the PRECE does not provide the efficient result. In our proposed concept PROCEOL first perform the noise removal process. The Noise removal means we are going to remove the noise presented in the given text. The noise removal performs the following operations such as the spell checking, abbreviations and word variants.

Experimental Testbed

Our experimental testbed is Lonely Planet corpus, consists of details about the countries, cities, cultures, organization, persons and etc. This dataset is taken from the http://www.lonelyplanet.com/destinations. We choose these data sets evaluating the performance of concept extraction for several considerations. First, these data sets enjoy different properties. Some of them are other than countries and cities details. Thus it is suitable for the PROCEOL technique.

6 5.5 5 4.5 4 3.5 Precision

V.2.

3

Comparison Process

Recall F1 Measure

2.5

In our experiments, we compare the proposed PROCEOL technique against state-of-the-art PRECE technique. The existing PRECE (Probabilistic Relational Concept Extraction) contains the term extraction and the Concept extraction techniques. Our proposed concept also uses the term extraction and concept extraction concept but with some changes. Normally the term extraction contains or performs the following operations such as the tokenization, parser, POS, Stemming, Lemmatization, and term weight. The first three processes Tokenization, Parser and POS are performed by using the GATE tool and the remaining operations are done using the Java. Tokenization is the process of separating or breaking the text into the words, symbols, phrases and other meaningful things. These are all called as the tokens. These tokens are separated by the normally whitespaces. The punctuation is not allowed in the tokenization process. Next step is the parser. It is preprocess of analyzing the sentence and phrase in terms of graphical constituents. Another is the POS (Parts Of Speech), this removes the noun, verb and adjectives present in the text. These above three are done with the help of the GATE tool. The Stemming is the process of removing continuous tense and past tense. Say for example, playing and played means remove the ing, ed and etc from the word and recover the word as play. This process is called as the stemming. Lemmatization is the process of grouping together the different inflected forms of a word. Say for example the car is matched as the cars. The next process is the term weight. In this stage we are going to calculate the weight for the each term. The stemming, lemmatization and the term weight process are all done using the java. The term extraction is the preprocessing stage. So it is common for both the PRECE (Probabilistic Relational Concept Extraction) and PROCEOL (Probabilistic Relation of Concept Extraction in Ontology Learning) methods. Another one important concept is the Concept

2 1.5 1 0.5 0 CPROCEOL

CPROCEOLdep

CPLSA

Fig. 3. Results of Lonely Planet Corpus

This is one of the additional feature is added in our proposed PROCEOL concept. The PRECE does not perform this kind of an operation. So, it does not provide the efficient and also which does not provide the accurate result. Our existing concept PRECE directly perform the concept identification process not perform this noise removal process. After the term extraction process is completed, we enter into the Noise removal process. First process is the spell checking. In this method we are going to analyze each and every thing present in our system. It finds out what are all the errors are presented in the text. So we get the accurate result for perform this task. Next method in the noise removal is the abbreviations. This means if we are going to search the thing in terms of an abbreviation means, for providing the best and correct result. Say for example if we are going to search the bank details means, if the bank name is SBI OR IOB means for providing this results also. Our existing PRECE does not perform this abbreviation searching operation. But our proposed system performs this task and provides the efficient result. The last method is the word variant, which means that the one word refers the different meaning, based upon the place or situation likewise synonyms. We also perform this word variant task also in our proposed PROCEOL concept. Say for example, we are going to search bank details, we give the word bank only but it retrieves the all the related word representing the word bank. These three processes are performed in our proposed concept.



723


TABLE I RESULT FOR OUR DATASET Technique

CPRECE 5.1726 0.3708 0.6919

Precision Recall F1 Measure

PRECE CPRECEdep 2.2708 0.1 0.1915

CPLSA 5.3415 0.6 1.0788

So it provides the accurate or best result. The next process is the co-occurrence analysis. This is the process of analyzing the co-occurrences present in the data. It is the process of identifying the similarities between the word pair and similarities meaning within word patterns. This is also one of the additional features of our proposed PROCEOL concept. After that we are enter into the concept identification. This concept identification uses the Markov logic network. The Markov logic network provides the secure result. The concept identification performs the two different processes such as the learning weight and another one is the inference. In existing concept the leaning weight presents one discriminative learning method only but in our proposed work we use the three different methods for presenting the learning weight such as the Discriminative learning, Generative learning, and Imitative learning. The Expectation Maximization process is also performed in our PROCEOL concept. In Expectation Maximization process the unknown truth values are handled. In our proposed concept for finding the probabilistic inference we are using MCMC (Markov Chain Monte Carlo) combination with the MC-SAT algorithm. Finally we perform the concept extraction process. This concept extraction performs the PLSA (Probabilistic Latent Semantic Analysis). In general, our technique is efficient and scalable for large applications. V.3.

PROCEOL CPROCEOLdep 2.4291 0.1 0.1920

CPROCEOL 5.4291 0.4291 0.7953

CPLSA 5.4291 0.6 1.0805

Finally we calculate the Term weight for the preprocessing process. This all process is covered by the PROCEOL techniques. CPROCEOLdep is the concept extraction using the PROCEOL technique using Markov Logic Network. This PROCEOL system extracts the concept without consider the syntactic dependencies between terms. CPLSA is the concept extraction using traditional Probabilistic Latent Semantic Analysis. Topic discovery and topic extraction process is performed using this Probabilistic Latent Semantic Analysis. Concepts are labeled with the term with highest probability given the particular topic. These three processes are calculated by the Precision, Recall and F1-Measures. [29] defined the calculation of the F1 measures. F1 measures is calculated as:

F1-measures=

2  Pr ecision  Re call Pr ecision  Re call

This is the basic formula for the F1-Measures. Here Precision and Recall is calculated by the formula of

Precision=

Recall=


Relevant documents  retrieved documents retrieved documents

Relevant documents  retrieved documents retrieved documents

Here the Relevant documents is said to be the Reference retrieval and the Retrieved documents is said to be the computed retrieval. So the final F1 measure is process of reference retrieval with the computed retrieval. Retrieved documents (i.e. computed retrieval) are the results of the gold standard, Table I shows the results of the precision, recall and F measures. The relevant documents (i.e. Reference retrieval) are the results of the concept of ontology learning.

This experiments aims to study the influence of concept extraction for the ontologies learning. Fig.4. shows the Precision, Recall and F1-Measures for the proposed technique. The PROCEOL technique was evaluated by comparing its output with a gold standard. For this purpose we use the data set of LonelyPlanet corpus for performing the concept extraction task. Here we evaluate three concept extraction techniques were used in order to extract the concepts from the LonelyPlanet corpus. This three extracted techniques are CPROCEOL, CPROCEOLdep, and CPLSA. CPROCEOL is the concept extraction using the PROCEOL technique. This PROCEOL technique is based on the Markov Logic Networks, which is performed by the Alchemy software packages. PROCEOL technique consists of the steps of Noise removal, Co-occurrence Analysis, Weight Learning and Inference calculations. Here the pre-processing steps like tokenization, parser, POS tags are performed by the GATE tool and also Stemming, Stop words, Lemmatization are performed by the java code.

VI.

Conclusion

This paper presented a framework of PROCEOL for concept extraction, which applies Markov Logic Network to learn ontologies with extraction. As a tradeoff between accuracy and efficiency, we first proposed deterministic PROCEOL technique using the Term Extraction and Concept Extraction. For implementing the PROCEOL technique, we use the dataset of lonely planet which contains the details of the tourism. First process of this concept extraction is term extraction. Term Extraction is used for extract the words.



724


This contains the process of Tokenization, Parser, Part of Speech analysis, stemming, Stop word, Lemmatization and finally the Term Weight. Second process is the concept extraction. In Concept extraction process, we presented the new tasks of noise removal and co-occurrence analysis in to the Markov Logic Networks. Most of the existing system only provides the concept identification process like learning weight with the inference and concept extraction using the PLSA method. But sometimes it makes the problems in extraction due the spell mistakes, problem of analysing abbreviations and issue of identifying word variants. This all problems are named as the Noise. Thus here we present the task of noise removal which is used for provide the process of spell checking, abbreviations analysis and word variants. And also we present another process of identifying the co-occurrence word. Co-occurrence word is the process of analyse the related or co-related words for the particular selected word. This co-occurrence is worked based on the wordnet tool. For implementing the pre-processing steps we use the GATE tool. That gate provides the accurate results for the pre-processing steps. Alchemy Packages also used for implement the concept extraction process. Alchemy packages are used for make the perfect MLN process. For implementing all this performance, we use the dataset of Lonely Planet. Lonely planet is the tourism dataset, contains the details of the country, organisation, cities and other region information. Lonely planet is easy help to understand the ontology process for users. Thus this PROCEOL method provides the effective results compared to the existing method. One of the drawbacks in our proposed system PROCEOL is that the system experiment with only one dataset Lonely Planet Tourism dataset, which contain text information in small size file and in English language. So, the system didn’t exercise with large size file with different file types and in different language. These are the limitation of the system. But the system get better F Score results compare to previous one. Because of the system are uses statistical relational learning technique which is already proved to give the better relationship between the terms. Another drawback of the system is, Concept Extraction is not sufficient for ontology learning, must learning Concept Hierarchy and establish Semantic relation in the Dataset.

Our idea is to provide the process of concept hierarchy extraction with the Integration of social data into the learning process. Integration of social data into the ontology learning is one of the best learning. So we decided to integrate the social data (like wikipedia). And also in future we have another idea to implement the semantic relations. We decided that to implement the semantics relations, I have idea of the using the association rule mining. And also the identification of axiom learning is important role of the ontology learning. So we need to implement the axiom learning also. For the entire step we finally implement the ontology population which is used to analyze the ontology populations. Thus we decided to implement the process for the future concepts.

References [1]

[2] [3]

[4]

[5]

[6]

[7]

[8]

[9]

[10]

[11]

[12]

VII.

Future Enhancement [13]

In this paper, we present the effective concept extraction technique named as PROCEOL. This provides the best noise removal with concept extraction results. Noise removal is the main process for remove the unwanted and error words in the data. Thus it performs the better process. But we make a decision to perform the more accuracy in concept extraction. In future, we decided to provide the more effective results in the concept extraction process.

[14]

[15]


Lucas Drumond and Rosario Girardi, “A Survey of Ontology Learning Procedures,” WONTO, volume 427 of CEUR Workshop Proceedings, CEUR-WS.org, vol. 427, 2008. Chris Biemann, “Ontology Learning from Text: A Survey of Methods,” LDV Forum, vol. 20, no.2, pp.75-93, 2005. Paul Buitelaar, Philipp Cimiano and Bernardo Magnini, “Ontology Learning from Text: An Overview,” Ontology Learning from Text: Methods, Evaluation and Applications, IOS Press, pp. 3-12, 2005. Cimiano P, Hotho A, Staab S, “Learning Concept Hierarchies from Text Corpora usingFormal Concept Analysis,” Journal of Artificial Intelligence Research, vol. 24, no.1, pp. 305–339, 2005. Cimiano P, Hotho A, Staab S, “Clustering Concept Hierarchies from Text,” Proceedings of the Conference on Lexical Resources and Evaluation (LREC), pp. 1721–1724, 2004. Lucas Drumond and Rosario Girardi, “An Experiment Using Markov Logic Networks to Extract Ontology Concepts from Text,” ACM Special Interest Group on Applied Computing, pp. 1354-1358, 2010. Daniel Lowd and Domingos.P, ”Efficient Weight Learning for markov logic networks”, Proceedings of the Eleventh European Conference on Principles and Practice of Knowledge Discovery in Databases, Warsaw, Poland, pp. 200-211, 2007. Tuyen N. Huynh and Raymond J. Mooney, ”Max-margin weight learning for Markov logic networks”, Proceedings of the European Conference on Machine Learning and Principles and Practice of Knowledge Discovery in Databases (ECML/PKDD09), Bled, Slovenia, pp.564-57, 2009. Parag Singla and Pedro Domingos, “Discriminative Training of Markov Logic Networks,” Proceedings of the 20th National Conference on Artificial Intelligence, vol. 2, pp. 868-873, 2005 Rohit Kate and Ray Mooney, “Probabilistic Abduction using Markov Logic Networks”, Proceedings of the IJCAI-09 Workshop on Plan, Activity, and Intent Recognition, 2009. Kaustubh Beedkar, Luciano Del Corro, Rainer Gemulla, “Fully Parallel Inference in Markov Logic Networks”, BTW, pp.205-224, 2013. Hassan Khosravi, ”Fast Parameter Learning for Markov Logic Networks Using Bayes Nets”, 22nd International Conference, Dubrovnik, Croatia, pp.102-115, September 17-19, 2012. Biba, M., Ferilli, S., and Esposito, F.,”Discriminative Structure Learning of Markov Logic Networks,” Proceedings of the 18th international conference on Inductive Logic Programming (ILP’08), Czech Republic. Springer-Verlag, pp. 59–76, 2008. Tuyen N. Huynh and Raymond J. Mooney, “Discriminative Structure and parameter learning for Markov Logic Networks”, Proceedings of the 25th International Conference on Machine Learning (ICML), New York,USA, Finland, pp. 416-423, July 2008. Shalini Ghosh, Natarajan Shankar, Sam Owre, “Machine Reading Using Markov Logic Networks for Collective Probabilistic


725


Inferences”, Proceedings of ECML-CoLISD, 2011. [16] Thomas Hofmann, “Probabilistic Latent Semantic Analysis,” Proceedings of 15th Conference on Uncertainty in Artificial Intelligence UAI’99, Stockholm, Sweden, pp.289-296., 1999. [17] Emhimed Salem Alatrish, “Comparison of Ontology Editors”, eRAF Journal on Computing, vol. 4, pp. 23 – 38, 2012. [18] Khondoker, M. Rahamatullah, Müller, Paul, “Comparing Ontology Development Tools Based on an Online Survey”, World Congress on Engineering 2010 (WCE 2010), London, UK, March 2010. [19] Mark Sanderson and Bruce Croft, “Deriving concept hierarchies from text”, SIGIR '99 Proceedings of the 22nd annual international ACM SIGIR conference on Research and development in information retrieval, pp. 206-213, 1999. [20] Maryam Hazman, Samhaa R. El-Beltagy and Ahmed Rafea, “Ontology Learning from Domain Specific Web Documents,” International Journal of Metadata, Semantics and Ontologies, vol. 4, no. 1/2, pp.24 – 33, 2009. [21] Zellig Sabbettai Harris, “Mathematical Structures in Language,” Interscience Publishers, p. 230, 1968. [22] Marti A.Hearst, “Automatic Acquisition of Hyponyms from Large Text Corpora,” COLING '92 Proceedings of the 14th conference on Computational linguistics, vol. 2, pp. 539-545, January 1992. [23] Karthikeyani.V. and Karthikeyan.K, “Migrate Web Documents into Web Data,” 3rd International conference on Electronics Computer Technology (ICECT), Kanyakumari, Tamil Nadu, vol. 5, pp. 249 – 253, 2011. [24] Karthikeyan.K and Dr.V.Karthikeyani, “Understanding text using Anaphora Resolution”, Internation conference on Pattern Recognition, Informatics and Mobile Engineering (PRIME), Salem, Tamil Nadu, pp- 346 – 350, 2013. [25] Hoifung Poon and Pedro Domingos, “Sound and Efficient Inference with Probabilistic and Deterministic Dependencies,” Proceedings of the 21st National Conference on Artificial intelligence, pp. 458-463, 2006. [26] Matthew Richardson.A and Pedro Domingos, “Markov Logic Networks”, Machine Learning, vol. 62, no.1-2, pp.107–136, 2006 [27] Thomas Hofmann, “Unsupervised Learning by Probabilistic Latent Semantic Analysis, Machine Learning, Kluwer Academic Publishers, vol. 42, no. 1-2, pp. 177-196, 2001. [28] Wong. W, Liu. W and Bennamoun. M., “Ontology Learning from Text: A Look Back and into the Future,” ACM Computing Surveys, vol. 44, no. 4, pp.30, August 2012. [29] Dellschaft.K, Staab.S, “On how to perform a gold standard based evaluation of ontology learning” Proceedings of ISWC-2006 International Semantic Web Conference, Athens, GA, 2006. [30] Nouri, Z., Nematbakhsh, M.A., Khayyambashi, M.R., Automated complementary association learning from web documents, (2009) International Review on Computers and Software (IRECOS), 4 (6), pp. 672-683. [31] D. N. Kanellopoulos, S. B. Kotsiantis, Semantic Web: A State of the Art Survey, (2007) International Review on Computers and Software (IRECOS), 2. (5), pp. 428 - 442. [32] Jabar, M.A., Khalefa, M.S., Abdullah, R.H., Abdullah, S., Metaanalysis of ontology software development process, (2014) International Review on Computers and Software (IRECOS), 9 (1), pp. 29-37.

Authors’ information Karthikeyan K. was born in Tamil Nadu, India in 13th February 1975. He is working as Assistant Professor in Department of Information Technology at NPR Arts & Science College, Natham, Tamil Nadu, India. He was received M.Sc degree in Information Technology from Bharathidasan University, Tamil Nadu, India in 2002, M.Phil degree from Madurai Kamraj University, Tamil Nadu, India in 2006. He has published and presented papers in 2 National and 2 International Conferences. He has 10 years teaching experience. His areas of Interests are Data Mining, Artificial Intelligence, Natural Language Processing, Data and Knowledge Engineering. Dr. Karthikeyani V. was born in Tamil Nadu, India in 11th June 1972. She is working as Assistant Professor in Department of Computer Science at Govt. Arts College, Rasipuram, Tamilnadu, India. She was awarded Doctoral degree from Periyar Universtiy, Salem, Tamilnadu, India. She has published 15 National and International Journals and presented several papers in International and National Conferences. She has 17 years of teaching experience. Her areas of interests are Image Processing, Computer Graphics, Multimedia, Data Mining and Web Mining. She is a life member in Computer Society of India (CSI), ISTE (Indian Society for Technical Education), ACM-CSTA (Computer Science Teacher Association) and various other International Computer Societies and organization for knowledge exchange and enhancement.



726


Artificial Fish Swarm Load Balancing and Job Migration Task with Overloading Detection in Cloud Computing Environments A. Mercy Gnana Rani1, A. Marimuthu2, A. Kavitha3 Abstract – Cloud computing is a widely used distributed computing model which offers a wide range of users with distributed access to scalable and virtualized hardware. Scheduling of tasks in cloud computing is a Nondeterministic Polynomial (NP) hard optimization problem and it has become an attractive research area in recent years. In task scheduling, the essential factor to be considered is the load balancing of non-preemptive independent tasks on Virtual Machines (VMs). In cloud computing system, the utilization of resources for the task and efficient detection of VM for task processing becomes difficult, as some VMs are heavily overloaded where as others are not heavily loaded. In order to overcome these problems, an efficient load balancing algorithm is used in this work. In this proposed system, initially overloading of VM are estimated based on the Hidden Semi Markov Model (HSMM) to maximize or minimize lifetime process, less latency and imbalanced degree for the clouds of different sizes. Artificial Fish is used to balance the task allocation for VM after overloaded is detected from HSMM. The proposed artificial fish swarm based algorithm effectively designs the priorities of tasks which in turn minimizes the waiting time of the tasks in the queue. The proposed approach shows that it achieves less unbalancing results and performs task faster with a reduction in waiting time of tasks in queue. Copyright © 2014 Praise Worthy Prize S.r.l. - All rights reserved.

Keywords: Load Balancing, Job Migration Task, And Overloading Detection, Cloud Computing, Hidden Semi Markov Model (HSMM), Artificial Fish Swarm Algorithm (AFSA)

The cloud computing services can be used from varied and extensive resources, rather than remote servers or local machines. In cloud computing, the distributed servers, offers the demanded services and resources to different clients in a network with the scalability and reliability of data center. The distributed computers provide on-demand services. Services may be of software resources (e.g. Software as a Service, SaaS) or physical resources (e.g. Platform as a Service, PaaS) or hardware/infrastructure (e.g. Hardware as a Service, HaaS or Infrastructure as a Service, IaaS) [2]. In a cloud computing paradigm, the random advent of tasks with random exploitation of CPU service time necessities can load a particular resource heavily, while the other resources are inactive or less loaded [3]. Hence, resource control or load balancing is the challenging issue in cloud computing. Load balancing is an approach to distribute workload across multiple systems, or other resources over the network links to attain optimal resource utilization, high throughput, and minimum response time with minimal overload. This research work is based on the simulation technique and it uses the cloud simulator, CloudSim [4]. The main aspect of these issues lies in the establishment of an effective load balancing algorithm. The load can be CPU load, memory capacity, and delay or network load [5]. Load Balancing is the process of distributing the loads among different nodes of a

Nomenclature

α N t t+1

Virtual Machine communication bandwidth ability Processing time maximum number of iterations number of tasks Current Fish (Task) Next Fist (task) position

I.

Introduction

Cloud computing involves several aspects which includes software, virtualization, distributed computing, networking, and web services. A cloud itself comprises of several components such as clients, data center and distributed servers with scalable, virtualized infrastructure over the internet [1]. It is possible to develop a new paradigm of ‘computing as a utility’ in the near future. In recent years, cloud computing has been widely used as an economic paradigm to acquire and manage IT resources. An institution requires weighing cost, advantages and issues of cloud computing before deploying it as an IT approach. The growth and development of recent technological innovations has resulted in the utilization of interconnected, multiple hosts rather than single high-speed processor which deserves cloud computing. Manuscript received and revised March 2014, accepted April 2014


727

A. Mercy Gnana Rani, A. Marimuthu, A. Kavitha

distributed system in order to enhance both resource utilization and job response time. This phenomenon eliminates the scenario in which some of the nodes are heavily loaded while other nodes are inactive. It is not always practical to maintain one or more idle services in order to accomplish the required demands. Tasks cannot be allotted to suitable servers and clients individually for efficient load balancing as the cloud framework is a very complex structure and its constituents are available throughout a widespread area. Thus, uncertainty may be attached while tasks are allotted. The essential aspects to be considered while developing such algorithm are load assessment, stability of different system, node interaction, CPU utilization, selection of nodes and the nature of work to be transferred [6]. This load could be represented in terms of CPU load, amount of memory used, delay or network load. Load balancing mechanisms can be generally classified as centralized or decentralized, dynamic or static, and periodic or non-periodic. Physical resources can be partitioned into a number of logical slices called Virtual Machines (VMs). The VM load balancing approaches are formulated to find out which VM is assigned to the next cloudlet [7]. In cloud computing scenarios, whenever a VM is heavily loaded with multiple jobs, these jobs have to be eliminated and submitted to the under loaded VMs of the same data center. In this scenario, when more than one task is eliminated from a heavy loaded VM and if there are more than one VM to process these tasks, these tasks have to be assigned to the VM such that there will be a balanced scheduling, i.e., no task should wait for a long time to get processed. Load balancing is carried out at VM intra-data center level. But, the overload detection of the host or virtual machine based on the time constraints are not performed at this level. This work focuses on the initial sub-problem of host overloads detection. When a host becomes overloaded, it directly affects the time constraints, as if the resource capacity is completely utilized. But, this would result in resource shortage and performance degradation of the system and the application. In this paper, a Hidden Semi Markov Model algorithm is used to evaluate the overload detection of host or VM. Load balancing is carried out using improved Artificial Fish Swarm Optimization for cloud computing environment. This proposed approach suggests that load balancing in cloud computing can be attained through modeling the fish behaviors such as preying, swarming, following with local search of fish individual for attaining the global optimum. The random and parallel search algorithm has significant potential to overcome local extrema and attaining the global extrema. This load balancing results have fast convergence speed.

II.

The major aspects that have to be noticed while developing such algorithms are the evaluation of load, comparison of load, stability of different system, performance of system, interaction between the nodes, characteristic features of work to be transferred and selection of nodes [7]. This load is evaluated in terms of CPU load, amount of memory used, delay or Network load. Majority of the scheduling approaches that have been presented earlier focused on maintaining the efficiency (load balancing) and equality for all tasks [8-10]. In this work, the authors reviewed the most appropriate research works made in the literature for job scheduling in cloud computing. Garg et al [11] presented the problem of increased energy consumption by data centers in cloud computing. A mathematical formulation is developed for energy efficiency based on several aspects like energy cost, CO2 emission rate, HPC workload and CPU power effectiveness. In this model, a near-optimal scheduling algorithm that uses heterogeneity across multiple data centers for a cloud provider was introduced. In [12], the author presented a task load balancing model in grid environment. It also provided the information about the system in a manner similar to the previous reference. This system transforms the Grid to tree structure independent of grid topological structure complication. The system will use this tree for load balancing. In [13], the author considered about three potentially feasible methods for load balancing in large scale cloud systems. In recent times, various nature inspired networking and computing models have established and applied for distributed methods to concentrate on the increasing scale and complexity in such systems. The honey-bee foraging in [14] is studied as a direct implementation of a natural phenomenon. After that, a distributed, biased random sampling method that preserve individual node loading near a global mean measure is examined. Finally, an algorithm for relating simile services by local rewiring is evaluated as a means of enhancing load balancing by active system restructuring. Each server processing a request from its queue and measures an income, which is similar to the quality that the bees show in their waggle dance. A load balancing model based on a tree representation of a grid is presented in [15], [16]. This load balancing approach has two main objectives namely minimization of the mean response time of tasks submitted to a grid and, the minimization of the communication costs during job transferring. This approach mainly deals with three layers of algorithms such as intra-site, intra-cluster and intra-grid. A significant advantage of this approach is that its choice for balancing the load is based on the current state of the system which helps in humanizing the overall performance of the system by migrate the load dynamically.

Literature Survey

A load balancing algorithm is dynamic in nature and does not take into account the earlier state of the system.



728


(

III. Proposed Methodology In cloud computing system, load balancing eliminates tasks from over loaded VMs and assigning them to under loaded VMs. Load balancing can affect the overall performance of a system. The task would be removed from host VM, when it reaches above the threshold of overload detection. The detection of overloaded task for each VM is important, so this work presents a semi hidden Markov model to estimate the state of VM group. Based on the detection of overloaded result, the decision is to balance the load, the scheduler should trigger the load balancing aspect using modified Artificial Fish Swarm Algorithm. Before identifying the overloaded condition for each VM, the problem of dynamic load balancing technique is initially determined which solves the problem of load imbalance between VMs. Load Balancing techniques considerably minimizes the makespan and response time. Initially, the condition of maximum lifetime is defined to complete the task. Makespan can be defined as the overall task completion time. The completion time of task on is denoted as . Hence, the makespan is defined as the following function: = max { | ∈ , = 1, … and ∈ , = 1,2, … ,

)=

(

)/

is defined as the VM utilization threshold differentiating the non-overload and overload stages of the host; represents the time, during which the host has been overloaded, which is a function of ; and is the total time, during which the host has been active. The following conditions need to be added to evaluate the job task when it exceed the overload condition: (

, ( (

)→ , ,

(6)

) ≤ )

(1)

=

×

+

+

, = 1, . .

=

) (8)

(9)

(2) Total length of tasks that are assigned to a VM is called a load:

By minimizing , Eq. (3) is obtained. From Eq. (2) and (3), equation (4) is obtained as follows: ≤

(

where processing element, denotes the number processors in , represents million instructions per second of all processors in and represents the communication bandwidth ability of : =

, = 1, …

(7)

where denotes the time when a VM migration has been initiated; represents the CPU utilization threshold defining the overload state of the host; ( )is the time, during which the host has , been overloaded, which is a function of and ; is the total time, during which the host has been active, which is also a function of and ; and is the ( ) value. limit on the maximum allowed It is necessary to exploit the mean time between VM movements started by the host overload detection algorithm, which can be attained by increasing each individual inter-migration time interval. Hence, the problem formulation is narrowed down to a single VM migration, i.e., the time span of a problem occurrence is from the end of a previous VM migration and to the end of the next. The capacity of:

Response time is the total time taken between submission of a request and the initial response that is produced. The minimization of waiting time is useful in improving the responsiveness of the VMs. Let ={ , ,… } be the set of virtual machines which should process n tasks represented by the set = { , … , } . All the machines are unrelated and parallel and are denoted as R in the model. The finishing time of a task is denoted by . The main aim is to minimize the makespan which can be denoted as . So this model is | | . Processing time of a task on virtual machine can be denoted as . Processing time of all tasks in a can be defined by Eq. (2): =

(5)

,

(3)

,

=

N T ,t  S VM i ,t 

(10)

The load of a VM can be evaluated as the number of tasks at time t on service queue of divided by the service rate of at time . With the aim of estimation the overloaded detection results for VM, a semi hidden markov models with ={ , ,… } be the set of virtual machines which should process n tasks

(4)

The threshold function of the CPU utilization load is defined as below:



729


represented by the set

= { ,…….,

} with (

). ( )= (

III.1. Hidden Semi Markov Model (HSMM)

(

=

| = = ( =

=

(

, |

, ≠

(15)

)=

= )

(16)

(17)

The completed state sequence complicates the likelihood function by an additional sum over all possible prolongations of the overload detection sequence ,… . It is given by:

(12) ( )=

( )

=

| )= ( = , = 1…. − 1 , , = ( ))

, =

For ∈ 1, … . Sojourn time is the occupancy of each and every state in SHMM to fulfill the task allocation process for VMs. The sojourn of the unobserved overloaded VM process length is denoted as ( ). It is between the limits + 1 to + in the state j. Every task in state j has to be in different state before and after the sojourn time. The upper bound of the time spent in state is denoted by . It is considered that the state occupancy distribution is focused on the finite set of time points {1 . . . , }, where may also raise up to the entire length of the observed sequence (in specific for parametric dwell time distributions). ( ) of the sojourn The entire process time time in state : ( )≔

( )=1

The state either overloads not below overload sequence remains in the last visited state from time − 1 to − 1 + , = 0, 1, . . . , R be the number of states in the SHMM the exit from the last visited state takes place at time − 1 + , which yields the completedate likelihood:

The occupancy of overload detection of VM distributions ( ) have to be assigned to each of the states by: = , ≠ )

= ) with ∑

The observation overload detection process is characterized by the conditional independence property:

( ) A HSMM comprises of pair of load overloading VM task of discrete-time VM processes { } and { } ∈ {0, … − 1}. The observed process { } is linked to the hidden, is defined as ( ) overload utilization process of each VM unobserved state processess final VM overload detection tasks { } by the conditional distribution depending on the state process. Here ={ ,… } and { } = { , … … . , } ∈ ( )} { Initial probabilities ≔ ( = ) with ∑ =1 The transition probabilities for the state . For each ≠ : ≔ ( = | ≠ , = ) (11) ∑ ℎ = 1 and =0

( )= ( ≠ | = 0 ,… − 2 | = ,

|

=

(

,

| )

(18)

,….

A task eliminated from overloaded VM has to determine the appropriate under load VMs. It has two possibilities, i.e., either it detects the VM set (Positive signal) or it may not determine the appropriate VM (negative signal). There could be more than one VM (a set of VMs for allocation) which can accept this task. Detecting the overload detection state of the VM group: ( ) is equal to the  If the VM load of ( ) estimation.  The System is overloaded.  Else.  The system is safe state perform load balancing using MAFSA.  Exit.

(13) Determining the overloaded group:

If the overload detection process starts in state time = 0, the following relation can be verified: (

≠ |

= ,

= 1 ,… ) =

( ) )

at  When the current workload of the VM group exceeds the maximum capacity of the group, then the group is overloaded. Load balancing is not possible in this scenario.  If ( ) > maximum capacity:

(14)

It is to be observed that the conditional independence of VM holds at each time step for a Markov chain, i.e., the current state at time is based on only on the last state at time − 1. HSMM is considered as a pair of stochastic processes{ , }. The final observed overload detection output process of is related to the semiMarkov chain by the observation (or emission) probabilities:

(

, ( (

, ,

)→ ) ≤ )

 Load balancing is not possible.  Else.



730


where Rand () produces random numbers between 0 and 1, Step is the step length to perform load balancing operation in the VM for every task, and is the optimizing load balancing parameter , n is the number of tasks in queue. Visual represents the visual distance from one task to another task for load balancing, and  is the  crowd factor (0 <  < 1). The functions of load  balancing task for every VM include the behaviors of the AF: AF_Prey, AF_Swarm, AF_Follow, AF_Move. Every task (fish) usually stays in the place with a best load balanced VM food, so simulate the behaviors of task allocation for VM based on this characteristic to find the best load balancing VM which is the basic idea of the AFSA. The basic behaviors of AF are defined [18]- [19] as follows for maximum. AF_Prey: This is a basic biological behavior that tends to the each task is allocated to best load balancing VM food; generally the fish (task) perceives the concentration of load balanced VM food in water to determine the movement by vision:

 Trigger load balancing using MAFSA. If the overload process is determined then load balancing is carried out using Modified Artificial Fish Swarm in which each fish acts as tasks for source overloaded detection VM from HSMM. III.2. Load Balancing of Tasks Using MAFSA Artificial Fish (AF) is a fabricated unit of true fish, which is used to formulate the analysis and explanation of the problem, and can be deployed through an animal ecology perception. AF acts as a task that realizes the virtual machine load, balancing the external perception as shown in Fig. 1. is the current state of the task that run on VM of an AF, visual represents the visual distance, and denotes the visual position of the current task at any moment. If the current task of the VM state at the visual position is better than the current state, it goes forward to the next task in the direction, and arrives at the state; otherwise, task allocation for load balancing continues an inspecting tour in the vision until it reaches maximum level of the VM load balancing. It does not need to travel throughout complex, which is helpful to find the global load balancing estimation results for task optimum by facilitating certain local optimum load balancing results for every VM with less overload task allocation with uncertainty load [17].

=

+

.

()

(21)

If < in the maximum problem that is overloaded data for each task then it goes to the forward to another VM to allocation task; otherwise, select a state randomly again to load balancing VM task and judge whether it satisfies the forward condition. If it cannot satisfy load balancing for given task to VM after maximum number of iterations completed by fish, it moves a step randomly to choose another VM. When the maximum number of iterations is small in AF_Prey, the AF (task) can work like a swim manner randomly, which makes it best local load balancing results: (

)

=

( )

+

.

()

(22)

AF_Swarm: The fish will assemble in groups of task that are naturally assign task to VM in the moving process, which is a kind of living habits to satisfy load balancing task for VM and avoid dangers stage in VM. Behavior description: Let be the AF current state of task in the VM process, be the center location of the task in the VM and be the number of its companions in the current neighborhood ( < ), n is total fish number. If > and > , which means that the companion center has more best load balancing results for VM and is not very crowded, it goes forward a step to the companion center:

Fig. 1. Vision concept of the Artificial Fish

Let is the current state of the task that run on VM } then process can = { , … , } and = { ,…, be expressed as follows:

(

=

+ =

. +

− |

− |

(), ∈ (0, ] .

()

(19) (20)

)

=

( )

+

−

( )

−

( )

.

()

(23)

Otherwise, executes the preying behavior. The crowd factor limits the length of the searching space in task for



731


load balancing, and more AF only cluster at the optimal area, which make sure that AF move to optimum in a wide field. AF_Follow: In the moving process of the task from one place to many places then find best load balancing by comparing the neighborhood partners will trail and reach the load balancing task quickly: (

)

( )

=

+

.

()

in demand in the course of dynamic provisioning or deprovisioning from clouds. Considering all these directly cannot use the cloud computing system. Experimenting new techniques or strategies in real cloud computing operations is not practically possible as such experiments will compromise the end users QoS requirements like security, cost, and speed. CloudSim [21]–[22] simulator is a widespread simulation structure that allows modeling, simulation and experimenting the cloud computing infrastructure and application services [21]. In this section, the performance of the proposed algorithm is analyzed based on the results of simulation using CloudSim. Then, the extended classes of CloudSim is used to simulate this algorithm. Fig. 2 illustrates the comparison of Makespan BL (Before Load) and After Load balancing using honey bee behavior inspired load balancing (AFLB- HBB)), After load balancing and job migration modified Artificial Fish Swarm Algorithm- (AFLB-MAFSA). The X-Saxis represents the number of tasks and the Y-axis represents the Makespan (task execution and completion time) in seconds, it shows that proposed AFLB-MAFSA are less execution time than the AFLB- HBB.

(24)

AF_Move: Fish (tasks) swim randomly in the water; in fact, they are seeking best load balancing for VM food or companions in larger ranges: (

)

=

( )

+

.

()

(25)

To increase probability value results of the load balancing by based on the fish behavior effort to include leaping behavior to AF. The AF’s leaping behavior is defined as follows. AF_Leap: If the objective functions of the load balancing that are OVMTS(vmu ) is less than L(θ) estimation for m and n iterations performed by fish for each VM. It chooses some fish (task) randomly in the entire fish (tasks) swarm, and set parameters randomly to the selected AF. is a parameter or a function that can make some fish have other abnormal actions (values), is a smaller constant [20] for each task: ( )− (

)

=

( )

( ) < . ()

+ .

(26)

The step direction is updated for next iteration of task: =

.

−

(27)

where is parameter to the maximum number of iterations and N be the number of tasks in, t is current fishes and t+ 1 next fish (task) position. This process is repeated until all the task is completed with less overloaded detection in a VM. The information the fishes (tasks) update the loaded on a VM, load on all VMs, number of tasks in each VM, the number of VMs in each VM. Load balanced VMs are not used in switching of tasks. Once the task switching is over, the balanced VMs are included into the load balanced VM set. Then this set has all the VMs, the load balancing is successful, i.e., all tasks are balanced.

IV.

Fig. 2. Comparison of makespan before and after load balancing using HBB and MAFSA

Fig. 3 illustrates the response time of VMs in seconds for AFLB-MAFSA and AFLB-HBB Algorithms.


A cloud computing system has to prevail over various hurdles such as network flow, load balancing of virtual machines, a federation of clouds, scalability and trust management etc. Clouds offer a set of services (software and hardware) on an unprecedented scale. Cloud Services have to handle the temporal variation

Fig. 3. Response time of VMs in seconds for AFLB-MAFSA and AFLB-HBB

The X-axis represents number of tasks and the Y-axis



732


represents time in seconds. It is evident that AFLBMAFSA is more efficient compared with other AFLBHBB: =

−

satisfaction. In this paper, an efficient HSMM algorithm for overload detection problem is proposed in the VM task scheduling process, once the VM task overloaded is detected, and then load balancing is applied through modified Artificial Fish Swarm Algorithm (AFSA) to overcome the problems of load balancing in VMs. Each fish’s act as a task and moves one to another VM once required load is balanced by system simultaneously allocated task are removed from existing tasks. The proposed algorithm improves the overall throughput of the processing and reducing the amount of time to complete task on VM. Thus, it reduces the response of time of VMs. Also, the precision analysis of overload detection is attained by HSMM model.

(28)

where & are the maximum and minimum among all VMs, is the average of VMs. This load balancing system reduces the degree of imbalance drastically. Fig. 4 shows the degree of imbalance between VMs before and after load balancing with AFLB-HBB and AFLB-MAFSA. The X-axis represents the number of tasks and the Y-axis represents the degree of imbalance. It is clearly evident that after load balancing with AFLBHBB and proposed AFLB-MAFSA, the degree of imbalance is greatly reduced. Figure 5 shows the comparison of the task migration between AFLB-HBB and AFLB-MAFSA. The X-axis represents the number of tasks and the Y-axis represents the number of task migrated. AFLB-MAFSA is more efficient and have a lesser number of tasks migrated when compared with AFLB-HBB.

References [1]

[2]

[3]

[4]

[5] [6]

[7]

[8] Fig. 4. Degree of imbalance between VMs before and after balancing [9]

[10]

[11] Fig. 5. Comparison of number of task migrations [12]

V.

Conclusion [13]

Load balancing is one of the major conflict in cloud computing. It is needed to dedicate the workload consistently among all the nodes in the network to achieve high resource utilization ratio and user

[14]

[15]


Klancnik,T., Blazic, B.J., Context-aware information broker for cloud computing, (2010) International Review on Computers and Software (IRECOS), 5 (1), pp. 52-58. Martin Randles, David Lamb, A. Taleb-Bendiab, A Comparative Study into Distributed Load Balancing Algorithms for Cloud Computing, 2010 IEEE 24th International Conference on Advanced Information Networking and Applications Workshops Livny, M.; Melman, M. (2011): Load Balancing in Homogeneous Broadcast Distributed Systems. Proceedings of the ACM Computer Network: Performance Symposium, pp. 47-55. CloudSim: A Framework for Modeling and Simulation of Cloud Computing Infrastructures and Services, The Cloud Computing and Distributed Systems (CLOUDS) Laboratory, University of Melbourne, (2011) available from: http://www.cloudbus.org/cloudsim/. A.Khiyaita and M.Zbakh, Load Balancing Cloud Computing: State of Art IEEE, pp. 106-109, May 2012. Alsaih, M.A., Latip, R., Abdullah, A., Subramaniam, S.K., A taxonomy of load balancing techniques in cloud computing, (2013) International Review on Computers and Software (IRECOS), 8 (1), pp. 64-76. Shu-Ching Wang, Kuo-Qin Yan, Wen-Pin Liao, Shun-Sheng Wang, Towards a Load Balancing in a Three-level Cloud Computing Network, 2010 IEEE, pp. 108-113. Li, G., Gu, Y., Xiong, H., Kong, J., Xu, S., A hybrid particle swarm algorithm for job shop scheduling problem, (2011) International Review on Computers and Software (IRECOS), 6 (6), pp. 1006-1012. Gupta, P.K. and N. Rakesh, 2010. Different job scheduling methodologies for web application and web server in a cloud computing environment. Proceedings of the 3rd International Conference on Emerging Trends in Engineering and Technology, Nov. 19-21, IEEE Xplore Press, Goa, pp: 569-572. DOI: 10.1109/ICETET.2010.24. Yang, B., X. Xu, F. Tan and D.H. Park, 2011. An utilitybased job scheduling algorithm for cloud computingconsidering reliability factor. Proceedings of the2011 International Conference on Cloud and Service Computing, Dec. 12-14, IEEE Xplore Press, Hong Kong, pp: 95-102. DOI:10.1109/CSC.2011.6138559 Garg, S.K., C.S. Yeo, A. Anandasivam and R. Buyya, 2009. Energy-efficient scheduling of HPC applications in cloud computing environments. Comput. Sci. Distributed, Parallel Cluster Computing. B. Yagoubi, M. Medebber, A load balancing model for grid environment, computer and information sciences, 2007. iscis 2007, in: 22nd International Symposium on, 7–9 Nov, 2007, pp. 1–7. H.D. Karatza, Job scheduling in heterogeneous distributed systems, Journal of Systems and Software 56 (1994) 203–212. Zhenhuan Gong, Prakash Ramaswamy, Xiaohui Gu, Xiaosong Ma(2009), “SigLM: Signature-Driven Load Management for Cloud Computing Infrastructures”IEEE. Arora M, S.k.Das and R.Biswas, 2002. “A decentralized


733


[16]

[17]

[18] [19]

[20]

[21]

[22]

scheduling and load balancing algorithm for heterogeneous grid environments”. In proc. of International conference on parallel processing workshops (ICPPW’02), Vancouver, British Columbia Canada, pp:499-505. HAN Xiangchun, PAN Xun. Distributed scheduling pattern for dynamic load balance in computing grid [J]. Computer engineering and Design,2007,28(12):2845-2847. M.Jian , Y.wang, F.Rubio, D.Yuan, ” Spread Spectrum code estimation by artificial fish swam algorithm block-coding and antenna selection, ” IEEE international symposium on intelligent signal processing (WISP), October 2007. X.L.Li, “A New Intelligent Optimization- Artificial Fish Swarm Algorithm, ” PhD thesis, Zhejiang University, China, June, 2003, M.Y.Jiang, D.F.Yuan, “Wavelet Threshold Optimization with Artificial Fish Swarm Algorithm, ” in Proc. of the IEEE International Conference on Neural Networks and Brain, (ICNN&B’2005), Beijing, China, 13-15, Oct. 2005, pp.569-572. C.Wang, C.Zhou, J.Wma , “an improved artificial fish-swarm algorithm and it `s application in feed-forward neural networks, ” Proceeding of fourth international conference on machine learning and cybernetics, August 2005, pp.18-21. R.N. Calheiros, R. Ranjan, A. Beloglazov, C.A.F. De Rose, R. Buyya, CloudSim: a toolkit for modeling and simulation of cloud computing environments and evaluation of resource provisioning algorithms, Software: Practice and Experience 41 (2011) 23–50, http://dx.doi.org/10.1002/spe.995. R.N. Calheiros, R. Ranjan, C.A.F.D. Rose, R. Buyya, CloudSim: a novel framework for modeling and simulation of cloud computing infrastructures and services, Computing Research Repository, vol. abs/0903.2525, 2009.

Dr. A. Kavitha received her M.C.A & M.Phil(Computer Science) from Bharathiar University, Tamil Nadu, India. She received her P.hD degree in Computer Science from V.M University, India. She spent nearly 18 years as a teaching faculty in Computer Science Departments at Various University Affiliated Colleges. She has published over 24 articles in reputed journals.

Author’s information 1

Assistant Professor in Department of Information Technology, Dr.SNS Rajalakshmi College of Arts and Science, Coimbatore-49. E-mail: [email protected] 2

Associate Professor, PG & Research Dept. of Computer Science, Govt. Arts College (Autonomous), Coimbatore-18. E-mail : [email protected] 3

Assistant Professor in Computer Science, Kongunadu Arts and Science College (Autonomous), Coimbatore-29. A. Mercy Gnana Rani received her M.C.A from Bharathidasan University, Tamil Nadu, India. She received her M.Phil degree in Computer Science from Mother Teresa Women’s University, Kodaikanal., India. She spent nearly 12 years as a teaching faculty in Computer Discipline Departments at Various University Affiliated Colleges. Currently she is working as Assistant Professor in Dr.SNS Rajalakshmi College Of Arts & Science(Autonomous),Coimbatore and pursuing PhD in the department of Computer science at Bharathiar University, Coimbatore. Her research area is Cloud Computing. Dr. A. Marimuthu received his M.C.A & M.Phil(Computer Science) degree from Bharathiar University, Coimbatore, India. He received his Ph.D. degree in Computer Science V.M University, India. He spent nearly 21 years as a teaching faculty in Computer Science Department. His main research is on Computer Networks, Software Engineering and Cloud Computing. He has published over 30 articles in reputed journals and is a reviewer in many International Journals. He acts as a chair person in many International Conferences.



734

1828-6011(201404)9:4;1-O Copyright © 2014 Praise Worthy Prize S.r.l. - All rights reserved