JOURNAL OF INFORMATION SYSTEMS ...

JOURNAL OF INFORMATION SYSTEMS & OPERATIONS MANAGEMENT

Vol. 7 No. 2 December 2013

EDITURA UNIVERSITARA Bucuresti

Foreword Welcome to the Journal of Information Systems & Operations Management (ISSN 1843-4711; IDB indexation: ProQuest, REPEC, QBE, EBSCO, COPERNICUS). This journal is an open access journal published two times a year by the Romanian-American University. The published articles focus on IT&C and belong to national and international researchers, professors who want to share their results of research, to share ideas, to speak about their expertise and Ph.D. students who want to improve their knowledge, to present their emerging doctoral research. Being a challenging and a favorable medium for scientific discussions, all the issues of the journal contain articles dealing with current issues from computer science, economics, management, IT&C, etc. Furthermore, JISOM encourages the cross-disciplinary research of national and international researchers and welcomes the contributions which give a special “touch and flavor” to the mentioned fields. Each article undergoes a double-blind review from an internationally and nationally recognized pool of reviewers. JISOM thanks all the authors who contributed to this journal by submitting their work to be published, and also thanks to all reviewers who helped and spared their valuable time in reviewing and evaluating the manuscripts. Last but not least, JISOM aims at being one of the distinguished journals in the mentioned fields. Looking forward to receiving your contributions, Best Wishes Virgil Chichernea, Ph.D. Editor-in-Chief


GENERAL MANAGER Professor Ovidiu Folcut EDITOR IN CHIEF Professor Virgil Chichernea EDITORIAL BOARD Academician Gheorghe Păun Academician Mircea Stelian Petrescu Professor Eduard Radaceanu Professor Ronald Carrier Professor Pauline Cushman Professor Ramon Mata-Toledo Professor Allan Berg Professor Kent Zimmerman Professor Traian Muntean

Lecturer Alexandru Tabusca

Romanian Academy Romanian Academy Romanian Technical Academy James Madison University, U.S.A. James Madison University, U.S.A. James Madison University, U.S.A. University of Dallas, U.S.A. James Madison University, U.S.A. Universite de la Mediterranee, Aix – Marseille II , FRANCE James Madison University, U.S.A. Louisiana Tech University, U.S.A. Romanian-American University Romanian-American University Academy of Economic Studies Academy of Economic Studies Academy of Economic Studies Academy of Economic Studies University “Politehnica” Bucharest University “Politehnica” Bucharest National Technical Defence University, Romania University “Babes-Bolyai” Cluj Napoca University “Babes-Bolyai” Cluj Napoca University “Politehnica” Bucharest University “Politehnica” Bucharest University “Politehnica” Bucharest The Technical University of Civil Engineering Bucharest Romanian-American University

Senior Staff Text Processing: Lecturer Gabriel Eugen Garais Assistant lecturer Mariana Coancă Assistant lecturer Dragos-Paul Pop

Romanian-American University Romanian-American University Romanian-American University

Associate. Professor Susan Kruc Associate Professor Mihaela Paun Professor Cornelia Botezatu Professor Victor Munteanu Professor Ion Ivan Professor Radu Şerban Professor Ion Smeureanu Professor Floarea Năstase Professor Sergiu Iliescu Professor Mircea Cirnu Professor Victor Patriciu Professor Stefan Ioan Nitchi Professor Lucia Rusu Professor Ion Bucur Associate Professor Costin Boiangiu Associate Professor Irina Fagarasanu Associate Professor Viorel Marinescu

JISOM journal details 2012

No. 1 2 3 4 5 6

Item Value Category 2010 (by CNCSIS) B+ CNCSIS Code 844 JOURNAL OF INFORMATION Complete title / IDB title SYSTEMS & OPERATIONS MANAGEMENT ISSN (print and/or electronic) 1843-4711 Frequency SEMESTRIAL Journal website (direct link to journal http://JISOM.RAU.RO section) ProQuest EBSCO

7

REPEC http://ideas.repec.org/s/rau/jisomg.ht IDB indexation (direct link to journal ml section / search interface) COPERNICUS http://journals.indexcopernicus.com/ karta.php?action=masterlist&id=514 7 QBE

Contact First name and last name Virgil CHICHERNEA, PhD Professor Phone +4-0729-140815 | +4-021-2029513 E-mail [email protected] [email protected]

ISSN: 1843-4711

The Proceedings of Journal ISOM Vol. 7 No. 2 CONTENTS Editorial VOTING-BASED IMAGE SEGMENTATION

211

THE IMPORTANCE OF DIGITAL MARKETING. AN EXPLORATORY STUDY TO FIND THE PERCEPTION AND EFFECTIVENESS OF DIGITAL MARKETING AMONGST THE MARKETING PROFESSIONALS IN PAKISTAN

221

IMAGE REPRESENTATION USING PHOTONS

229

DATABASE DYNAMIC MANAGEMENT PLATFORM (DBDMS) IN OPERATIVE SOFTWARE SYSTEMS

239

TEXT LINE SEGMENTATION IN HANDWRITTEN DOCUMENTS BASED ON DYNAMIC WEIGHTS

247

Ion Ivan Alin Zamfiroiu

M-TOURISM EDUCATION FOR FUTURE QUALITY MANAGEMENT

264

Irina Mocanu Tatiana Cristea

HAND GESTURES RECOGNITION USING TIME DELAY NETWORKS

272

CIRCULAR CONVOLUTION AND FOURIER DISCRETE TRANSFORMATION

280

MANAGEMENT OF INNOVATION IN THE MODERN KAZAKHSTAN: DEVELOPMENT PRIORITIES OF SCIENCE, TECHNOLOGY AND INNOVATION

288

2:1 UPSAMPLING-DOWNSAMPLING IMAGE RECONSTRUCTION SYSTEM

294

Andreea-Mihaela Pintilie Costin-Anton Boiangiu

STUDY OF NEUROBIOLOGICAL IMAGES BASED ON ONTOLOGIES USED IN SUPERRESOLUTION ALGORITHMS

300

Crișan Daniela Alexandra Stănică Justina Lavinia

A FUZZY COGNITIVE MAP FOR HOUSING DOMAIN

309

Costin-Anton Boiangiu Radu Ioanitescu Fawad Khan Professor Dr Kamran Siddiqui

Andreea-Mihaela Pintilie Mihai Zaharescu Ion Bucur Virgil Chichernea Dragos-Paul Pop Costin-Anton Boiangiu Mihai Cristian Tanase Radu Ioanitescu

Mircea Ion Cîrnu Rauan Danabayeva

Mihai Cristian Tănase Mihai Zaharescu Ion Bucur

Cristina Coculescu

POSSIBILITIES OF DYNAMIC SYSTEMS SIMULATION

319

Alexandru Tăbușcă

HTML5 – AUGMENTED REALITY, A NEW ALLIANCE AGAINST THE OLD WEB EMPIRE?

325

Gabriel Eugen Garais

CASE STUDY ON HIGHLIGHTING QUALITY CHARACTERISTICS OF MAINTAINABLE WEB APPLICATIONS

333

Camelia M. Gheorghe Mihai Sebea

MANAGING TECHNOLOGICAL CHANGE IN INTERNATIONAL TOURISM BUSINESS

343

VERIFIABLE SECRET SHARING SCHEME BASED ON INTEGER REPRESENTATION

350

THE IMPACT OF 3D PRINTING TECHNOLOGY ON THE SOCIETY AND ECONOMY

360

CONSIDERATIONS REGARDING THE INTERNET PURCHASES BY INDIVIDUALS IN ROMANIA AND EUROPE

371

AN OVERVIEW OF DOCUMENT IMAGE ANALYSIS SYSTEMS

378

ASSISTIVE I.T. FOR VISUALLY IMPAIRED PEOPLE

391

PROJECT MANAGEMENT DATA IN INNOVATION ORIENTED SOFTWARE DEVELOPMENT

404

MODELING AND OPTIMIZING THE BUSINESS PROCESSES USING MICROSOFT OFFICE EXCEL

414

A COMPARISON OF SOME NEW METHODS FOR SOLVING ALGEBRAIC EQUATIONS

422

Qassim Al Mahmoud Alexandru Pîrjan Dana-Mihaela Petroşanu Marian Zaharia Daniela Enachescu Andrei Tigora Oana Bălan Alin Moldoveanu Florica Moldoveanu Anca Morar Victor Asavei Mihai Liviu Despa

Mihălcescu Cezar Sion Beatrice Anda Elena Olteanu Mircea Ion Cîrnu


VOTING-BASED IMAGE SEGMENTATION Costin-Anton Boiangiu1 Radu Ioanitescu2 ABSTRACT When it comes to image segmentation, there is no single technique that can provide the best possible result for any type of image. Therefore, based on different approaches, numerous algorithms have been developed so far and each has its upsides and downsides, depending on the input data. This paper proposes a voting method that tries to merge different results of some well-known image segmentation algorithms into a relevant output, aimed to be, as frequently as possible, better than any of the independent ones previously computed. KEYWORDS: image segmentation, machine vision, voting, cluster identification, image processing 1. INTRODUCTION In computer vision, segmentation refers to the process of partitioning an image into multiple sets of pixels based on similarities. This is useful when it comes to extracting specific data, for example, because it makes the image easier to analyze. Segmentation is used for locating particular elements or boundaries (lines, curves etc.) in the image. The result of the image segmentation process is a set of collections of pixels – or clusters of pixels – that covers the entire image. Each of the pixels in one cluster is similar to the others in the same cluster regarding a certain characteristic or property, such as color or intensity, and based on the same criteria, adjacent clusters are significantly different. This paper presents a method for segmenting an image using a voting technique that merges independent results of several known algorithms into a single final output, aimed to be, in as many cases as possible, better than any of the others in particular. 2. RELATED WORK The number of image segmentation techniques has grown rapidly in the last decades and more and more fields benefit from its advantages. Having a wide range of applications, segmentation techniques tend to be specific and cannot be applied in a wide range of cases. This creates a burden on the user that has to select the right segmentation algorithm for a given task which is quite hard considering how complex the algorithms are and their sometimes chaotic behavior on some specific sample inputs. One such image segmentation application example would be for medical data sets [1], [2], which has to address problems like low image contrast, quality of input and high variance of input cases. A large number of papers [3-12] have been published with the purpose of evaluating the state of the art in the image segmentation domain and trying to provide explanations on what are the best choices for a given of image characteristics. However, this documentation 1 Associate Professor PhD Eng., ”Politehnica” University of Bucharest Romania, 060042 Bucharest, [email protected] 2 Engineer, European Pattent Office (EPO) Germany, Bayerstrasse 34, Munich, [email protected]

211


includes only a limited number of segmentation algorithms, which are hard to evaluate and fine-tune for every image specific. In [13] a generic evaluation framework for image segmentation is presented. Ana Fred proposes [14] a majority voting combination of clustering algorithms. In short, a set of partitions is obtained by running multiple clustering algorithms and then the outputs are merged into a single result. The main idea behind majority voting is that the judgment of a group is superior to that of the individuals. Other papers like [15] try to use voting for removal of noise and other inherent artifacts that results from the segmentation process with efforts going as well into providing specific mathematical models, methods [16] and probabilistic analysis in order to evaluate the relative quality of segmentation [17]. The next section briefly presents the technique described in [14]. The results of the clustering algorithms are mapped into a matrix, called the co-association matrix, where each pair (i, j) represents the number of times the elements (pixels, in our case) i and j were found in the same cluster. A pseudo-code for the algorithm is: 1. Input: a set of A partitions obtained from running different A clustering algorithms on the initial data set of N samples 2. Output: a partition of the initial data set 3. Initialization: Set the co-association matrix to a null N x N matrix 4. for each clustering algorithm update the co-association matrix: a. for i = 1 to A, let P be the ith partition from the input set corresponding to the result of the ith algorithm b. for every sample pair (i, j) that belong to the same cluster in P set coassoc(i, j) = coassoc(i, j) + 1/A 5. obtain the output partition by thresholding on the matrix values: a. for each sample pair (i, j) that has coassoc(i, j) > threshold, combine i and j in the same cluster and also combine their belonging clusters, if necessary b. for the remaining samples form an one element cluster 6. return the result. This served as a starting point for the implementation of the voting image segmentation algorithm proposed by this paper. 3. AUXILIARY ALGORITHMS Different segmentation algorithms have been designed having qualities and shortcomings and articles like [3], [7], [9] have been dedicated to studying the different trade-offs that have to be made. This section briefly describes the individual image segmentation algorithms used in this paper to demonstrate the proposed concepts, with an emphasis on their upsides and downsides in terms of over/under segmentation i.e. computing too many/few segments as output for a given input image than what is ideally expected. An in-depth analysis of each is not made, as this is not the purpose of the paper. A series of representative algorithms from different categories in the image segmentation field are considered in order not to restrain the potential application domain. The paper 212


makes use of the following segmentation techniques: histogram based segmentation, region growing based on neighboring pixels, segmentation using graphs, watershed transformation and mean shift segmentation. Histogram based segmentation This algorithm focuses exclusively on pixel intensity in an image. The general idea behind it is the use of local minima and maxima as the boundaries for a cluster of pixels. Thus, the histogram of the image is required.

Figure 1. Histogram of a grayscale image

Fig. 1 displays a typical histogram of an 8-BPP (256 levels) grayscale image. The “peaks” represent the local maxima and the lower values represent the local minima. Between each adjacent minima/maxima or maxima/minima pair lies a cluster of pixels. This technique is described in detail in [18], [19] and is known to give good results when the foreground of the image is clearly delimited from its background. Because the pixel distribution in the image is not taken into account, using this method on other types of images does not lead to very relevant results. A histogram based segmentation algorithm is generally considered to over-segment the input image. Region growing based segmentation This is a simple concept, based on the assumption that an object does not generally change color attributes in an abrupt manner. Therefore, neighboring pixels (that belong to the object in question) should have relatively close intensity values. Thus, starting with a random pixel and “growing” the region around it, pixels with similar characteristics can be found based on a homogeneity criterion – close intensity values – and clustered in the same collection. Optimized implementations of the region growing segmentation algorithms are the seeded region growing described in [20] and the adaptive region growing described in [21]. Both algorithms produce satisfactory results on most images, but generally have a tendency towards over-segmentation. Graph based segmentation The technique used is based on the “minimal spanning tree” concept and is described in detail in [22], [23]. In short, a graph is built with nodes as pixels and edges that have an associated weight which measures the dissimilarity of the connected pixels. The segmentation is obtained by partitioning the graph into connected components so that the edges between the nodes in the same component have higher weights (the higher the weight, the lower the dissimilarity). 213


The algorithm is relatively accurate, under-segmenting the input image in some cases. Watershed based segmentation The Vincent-Soille algorithm is the classic approach when it comes to watershed based segmentation. A detailed description of this technique can be found in [24]. A prerequisite for watershed transformation is blurring the input image first, so applying a Gaussian filter to it is a recommended pre-processing step. Even so, the algorithm generally tends to strongly over-segment the image. Mean shift based segmentation The theoretical background behind this approach can be found in [25] and an actual implementation is given in [26]. Most of the times the results obtained with this algorithm are accurate, with a slight tendency for under-segmentation. 4. THE VOTING ALGORITHM The technique proposed is a variation of the more general algorithm based on majority voting clustering presented in the second section of this paper. Our algorithm mainly follows the same processing steps, with two major differences however: not all voting algorithms have the same “power of decision” and the actual implementation takes into consideration – and avoids – using too much of the system’s resources. Coping with memory requirements When images are used as input data, each pixel is a valid sample, so the total number of samples is given by the width of image times its height. Considering this and the fact that the co-association matrix is defined as an N x N matrix (where N is the number of samples), the result is that a very large structure will be needed to store the data. This could easily result in exceeding the addressing capabilities of a 32 bit operating system. To avoid such a scenario, a square-shaped sliding window across the image is used and only the pixels inside it are processed at a given time. The co-association matrix is thus dependent on the size of that window – adjustable by the user – instead of the image’s entire size. After the matrix is filled with data, the second step of the algorithm takes place, where pixels are being merged into clusters. This is done by assigning a label to those pixels – all the pixels in a cluster will be given the same label, which is different from labels associated with pixels in adjacent clusters. Once this is complete, the window slides forward and the process is repeated. Another aspect is the increment used to slide the window. If it were to slide by its entire size, then different clusters would be obtained in each region of the image the window overlaps, although a certain cluster could easily span across multiple such regions. A workaround to this is not to slide the window by its entire size, but rather by its size minus one row of pixels – both horizontally and vertically, where applicable. This way an overlap with previously considered regions is obtained and the erroneous splitting of a pixel set is avoided, because the first row and/or column of pixels are already labeled and therefore newly considered pixels can be added to previously found clusters. 214


Figure 2. The sliding window technique

Fig. 2 displays the process described above. The sliding window is showed in two consecutive iterations – the black thick frame and the grey thick frame. Notice the overlapping area between the two windows, which allows the grey pixels inside the second one to be merged into the cluster found by the first one. Weighted voting Because each of the segmentation algorithms used had its “strengths and weaknesses”, a method of selecting the better results was needed. The main choice criterion was the one used for analysis in the third section of our paper: the over/under segmentation tendency of an algorithm. Over-segmentation implies that different clusters contain pixels with the same characteristics and thus can be combined in a larger one. Contrary, under-segmentation represents the case when clusters span over regions with different characteristics that should be split. The assumption used was that if an algorithm is known to strongly over-segment an image and determines that two pixels are in the same cluster, then this is highly probable to be true. Analogous, if an algorithm is known to strongly under-segment an image and determines that two pixels are not in the same cluster then this is highly probable to be true as well. Considering all aforementioned assumptions, a performance evaluation parameter was introduced for each considered segmentation algorithm - called the over-segmentation factor. This parameter ranges from 0 to 1, where 0 represents the factor for the most extreme under-segmentation algorithm which would consider all pixels to be in one cluster, 1 the factor for the most extreme over-segmentation algorithm which will consider each pixel to have its own cluster, and 0.5 being the optimal segmentation algorithm which depends on the image characteristics. Based on the dynamic analysis of the output results of the independent segmentation algorithms used, an over-segmentation factor is associated with each as follows: a low one (user defined) for the algorithm that produces the fewest clusters, a high one (also user defined) for the algorithm that produces the most clusters and a variable one to the other algorithms (computed using linear interpolation based on the number of clusters they produce). This is a preprocessing step. Next comes the filling of the co-association matrix: when an algorithm finds two pixels that are/aren’t in the same cluster, a value proportional to its over-segmentation factor is added/subtracted to/from the corresponding matrix element. Thus the “weight” of each segmentation algorithm is determined in a manner that encourages its decision if the probability of it being true is high.

215


The second step of the original approach remains the same – mainly clusters are determined using a threshold value on the matrix. This is also user-defined. Because the lowest and the highest values of the over-segmentation factor, as well as the threshold value used in the cluster creation process are user-defined, the accuracy of the final result is dependent on how well these parameters are chosen. Complexity analysis For the first step of the algorithm, the complexity is O(N2). The second step introduces the cluster merging, which is done in O(logN) using a tree forest with path reduction. Therefore the total complexity of the algorithm is O(N2logN). It should be noted that the complexity depends on the image pixel size, so N represents the image width times its height. 5. EXPERIMENTAL RESULTS The algorithm performed well in general, giving good results on several test images considered. The system used for testing had an Intel® Core™2 Duo 2.4 GHz Processor and 2 GB RAM. The first test considered was the classic “Lena”, a perfect image for its natural color balance and great range of frequencies encountered across different areas:

Figure 3. Lena input image

The results obtained when running the algorithm set on the image shown in Fig. 3 are presented in Fig. 4 (the outputs for each independent algorithm, as well as the final output of the voting algorithm).

Figure 4. Lena segmentation processing results. From left to right and top to bottom: Region segmentation, Histogram segmentation, Watershed segmentation, Mean shift segmentation, Graph segmentation and (the final result) the Voting segmentation

216


As can be seen in Fig. 4, the region growing, histogram and watershed transformation based algorithms tend to over-segment the input image – especially watershed, which therefore will have the highest over-segmentation factor associated. On the other hand, it can be observed that the mean shift and graph segmentation methods tend to under-segment the image. In this case, the graph based segmentation algorithm will have the lowest oversegmentation factor associated. The voting based image algorithm obtains a balance between the over and undersegmentation characteristics of the individual algorithms. The current application is built with extensibility in mind, taking a framework approach where adding a new algorithm is as simple as implementing an interface. The framework is efficient because it runs all configured input segmentation algorithms in parallel. The approach presented in this paper is an important part of a bigger project: a complete, modular, fully automatic content conversion system developed for educational purposes. Once finished, the system will allow large batches of documents to be processed fully automatically and as a result more complex algorithms like [27] will be employed to provide input segmentations. The voting process tuning will also benefit from the availability of more diverse test data, allowing every input algorithm to adjust over/under segmentation by using its running history. Here are some other test images and the associated results for each of them (only the final voting based result is shown):

Figure 5. Cameraman segmentation processing results

Figure 6. Livingroom segmentation processing results

Figure 7. Mandril segmentation processing results

Figure 8. Peppers segmentation processing results

217


Figure 9. Pirate segmentation processing results Figure 10. Text segmentation processing results

Figure 11. Walkbridge segmentation processing results

Figure 12. Blonde woman segmentation processing results

Figure 13. Museum segmentation processing results

Figure 14. Question mark segmentation processing results

Figure 15. Map segmentation processing results

In the following table are presented the times obtained (for the voting algorithm only) on the considered testing system: Image size Size (pixels) Lena 512 x 512 Cameraman 512 x 512 Livingroom 512 x 512 Mandril 512 x 512 Peppers 512 x 512 Pirate 512 x 512 Text 975 x 821 Walkbridge 512 x 512 Blonde woman 512 x 512 Question mark 300 x 375 Museum 584 x 504 Map 512x512 Input image

Color depth (bits) 24 8 8 24 8 24 8 8 8 24 24 24

Table 1 Output result analysis

218

Run time (seconds) 22.03 19.14 13.26 13.82 14.44 14.48 50.40 12.62 15.20 11.21 23.48 16.16


6. CONCLUSIONS AND FUTURE WORK The aim of this work is to propose a weighted voting algorithm that segments an image based on the results of other well-known segmentation algorithms. The algorithm copes with large memory requirements and also takes into account the tendency of the individual considered segmentation algorithms to over/under segment and has encouraging higher accuracy results. Using the proposed algorithm, the results obtained were promising for most images tested. It can be therefore concluded that the idea of using several segmentation algorithms and merging their output to obtain a relevant result is viable. Another algorithm optimization approach could use a feedback loop by considering the results of some segmentation algorithms to adjust the input of the other segmentation algorithms from the framework. By knowing the general characteristics of a segmentation algorithm and its results on an image, some meaningful information about the processed image can thus be derived and looped back either into the voting process or the individual segmentation parameters. As future work there are two improvements to be taken into consideration. The first is adding an extra post-processing step for eliminating all the small clusters which may appear as being noise in the final result. The second is dynamically adapting the threshold parameter used for the merging/splitting of clusters, thus ensuring a higher chance of getting improved output results for a very varied range of input images. 7. ACKNOWLEDGMENT The authors would like to thank Constantin Manoila and Lucian-Ilie Calin, for their great ideas, support and assistance with this paper. 8. REFERENCES [1] P. Herghelescu, M. Gavrilescu, V. Manta, "Visualization of Segmented Structures in 3D Multimodal Medical Data Sets", Advances in Electrical and Computer Engineering, vol. 11, no. 3, 2011, pp. 99 - 104. [2] L. D. Pham, X. Chenyang and J. L. Prince, "Current Methods in Medical Image Segmentation". Annual Review of Biomedical Engineering, vol 2, pp. 315–337, 2000. [3] S. K. Pal, N. R. Pal, “A Review on Image Segmentation Techniques”, Pattern Recognition, Vol. 26, No. 9, pp. 1277-1294, 1993. [4] Fu K.S.,Mui J.K., "A survey on image segmentation", Elsevier Pattern Recognition journal, Volume 13, Issue 1, pp. 3–16, 1981. [5] R.M. Haralick, L.G. Shapiro, "Image Segmentation Techniques", CVGIP: Graphical Models and Image Processing vol. 29, pp. 100-132, 1985 [6] A. Hoover, "An experimental comparison of range image segmentation algorithms", IEEE Transactions on Pattern Analysis and Machine Intelligence, Volume 18, Issue 7, pp. 673-689. [7] A. A. Aly, S. B. Deris, N. Zaki, Research review for digital image segmentation techniques, International Journal of Computer Science & Information Technology (IJCSIT) vol. 3, no. 5, oct 2011 [8] Y.J. Zhang, “A survey on evaluation methods for image segmentation,” Pattern Recognition, vol. 29, no. 8, pp. 1335C1346, 1996.

219


[9] H. Zhang, J. E. Fritts, S. A. Goldman, “Image segmentation evaluation: a survey of unsupervised methods”, Computer Vision and Image Understanding vol. 110, Elsevier Science Inc., New York, pp. 260-280, 2008. [10] J. Freixenet, X. Munoz, D. Raba, J. Martí and X. Cufí, "Yet Another Survey on Image Segmentation: Region and Boundary Information Integration", Lecture Notes in Computer Science, vol. 2352/2002, pp. 21-25, 2002. [11] K. McGuinness, N. E. O’Connor, “A comparative evaluation of interactive segmentation algorithms”, Pattern Recognition vol. 43, Elsevier Science Inc., New York, pp. 434-444, 2010. [12] S. Chabrier, H. Laurent, B. Emile, C. Rosenburger, and P. Marche, “A comparative study of supervised evaluation criteria for image segmentation”, EUSIPCO, pp. 1143-1146, 2004. [13] J. S. Cardoso and L. Corte-Real, “Toward a Generic Evaluation of Image Segmentation”, IEEE Transactions on Image Processing, No. 11, pp. 1773-1782. [14] A. L. N. Fred, Finding consistent clusters in data partitions, Lecture Notes in Computer Science vol. 2096, Springer-Verlag, London, pages 309-318, 2001. [15] P. Jonghyun, T. K. Nguyen and L. Gueesang, "Noise Removal and Restoration Using VotingBased Analysis and Image Segmentation Based on Statistical Models", Energy minimization methods in computer vision and pattern recognition, Lecture Notes in Computer Science, vol 4679/2007, 242-252, 2007. [16] P. Correia and F. Pereira, “Objective evaluation of relative segmentation quality”, Image processing, vol 1, pp. 308-311, 2000. [17] K. Vincken, A. Koster and M. Viergever: Probabilistic multiscale image segmentation, IEEE Transactions on Pattern Analysis and Machine Intelligence, 19:2, pp. 109–120, 1997. [18] L. G. Shapiro, G. C. Stockman (2001): “Computer Vision”, New Jersey, Prentice-Hall, ISBN 013-030796-3, pp 279-325, 2001. [19] R. Ohlander, P. Keith, R. D. Raj, "Picture Segmentation Using a Recursive Region Splitting Method", Computer Graphics and Image Processing, vol. 8, no. 3, pp. 313–333, 1978. [20] R. Adam, L. Bischo: "Seeded Region Growing", IEEE Transactions On Pattern Analysis Machine Intelligence, vol. 16, no. 6, pp. 641-647, 1994. [21] Y. L. Chang, X. Li: “Adaptive Image Region-Growing, IEEE Transactions on Image Processing”, vol. 3, no. 6, pp. 868-872, 1994. [22] P. F. Felzenswalb, D. P. Huttenlocher, ”Efficient graph-based image segmentation”, International Journal of Computer Vision vol. 59, Kluwer Academic Publishers, Hingham, pp. 167-181, 2004. [23] J. Shi, J. Malik, “Normalized Cuts and Image Segmentation”. IEEE Trans. Pattern Analysis and Machine Intelligence, Vol. 22, No. 8, pp. 888-905, 2000. [24] J. Roerdink, A. Meijster, “The watershed transform: definitions, algorithms and parallelization strategies”, Fundamenta Informaticae vol. 41, IOS Press, Amsterdam, pp. 187-288, 2000. [25] D. Comaniciu, P. Meer, “Mean shift: a robust approach towards feature space analysis”, IEEE Transactions on Pattern Analysis and Machine Intelligence, vol. 24, IEEE Computer Society, Washington DC, pp. 603-619, 2002. [26] C. M. Christoudias, B. Georgescu, P. Meer: “Synergism in low level vision”. 16th International Conference on Pattern Recognition., Quebec City, Canada, August 2002, vol. IV, 150-155, 2002. [27] E. Ganea, D. D. Burdescu, M. Brezovan, "New Method to Detect Salient Objects in Image Segmentation using Hypergraph Structure", Advances in Electrical and Computer Engineering, vol. 11, no 4, pp. 111-116, 2011

220


THE IMPORTANCE OF DIGITAL MARKETING. AN EXPLORATORY STUDY TO FIND THE PERCEPTION AND EFFECTIVENESS OF DIGITAL MARKETING AMONGST THE MARKETING PROFESSIONALS IN PAKISTAN Fawad Khan1 Professor Dr Kamran Siddiqui2 ABSTRACT The purpose of this exploratory research is to present the perceptions towards Digital Marketing in Pakistan. This issue has rarely been addressed by the academicians and researchers in Pakistan and elsewhere. This study used digital marketing parameters to measure the awareness and effectiveness of digital marketing among marketing professionals in Pakistan. 200 marketing professionals participated in this academic exercise. Data was analyzed in many ways, a) through descriptive statistics b) summarizing the data using factor analysis. Four major perception groups were emerged from the analysis i.e., a) Skeptical b) Enthusiast c) Utilitarian and d) Parsimonious. The result suggests that professionals in Pakistan are more skeptical towards digital marketing tools and concepts. They do not fully understand the benefits of digital marketing in terms of growth and cost effectiveness. Finally, the limitations of the studies and findings are presented in study. Key words: SEO, Google Analytics, META tags, Blogs 1. INTRODUCTION There are not many studies conducted in Pakistan in the area of digital marketing. This concept is rapidly emerging as a new concept which is aggressively adopted internationally for marketing success. In today’s time, social media channels such as Face book, Twitter, Google and other social media firms have successfully transformed the attitudes and perceptions of consumers and in the end helped revolutionized many businesses. This was done through measurable vast network of customers with trustworthy data with real-time feedback of customer experiences. It is much more convenient for businesses to conduct surveys online with a purpose to get relevant information from targeted groups and analyzing the results based on their responses. Potential customers can look for reviews and recommendations to make informed decisions about buying a product or using the service. On the other hand, businesses can use the exercise to take action on relevant feedback from customers in meeting their needs more accurately. Change is constant and with time new ideas are accepted and adopted. In order to make the decision to understand the advantage of online marketing, advantages must be highlighted for industry players to realize its power.

1 DHA Suffa University,Phase VII (Ext), DHA, Karachi-75500, PAKISTAN., e-mail: [email protected], Tel: +923022914846, Fax: +9221-35244855 2 DHA Suffa University, Phase VII (Ext), DHA, Karachi-75500, PAKISTAN., e-mail: [email protected], Tel: +9221-35244865, Fax: +9221-35244855

221


2. LITERATURE REVIEW Literature Review The purpose of doing research in the area of digital marketing is because it seem huge, intimidating and foreign. Businesses are looking for clearer picture to start but do not know where and how to start doing digital marketing. In today’s time, social media channels such as Face book, Twitter, Google and other social media firms have successfully transformed the attitudes and perceptions of consumers and in the end helped revolutionized many businesses. This was done through measurable vast network of customers with trustworthy data with real-time feedback of customer experiences. It is much more convenient for businesses to conduct surveys online with a purpose to get relevant information from targeted groups and analyzing the results based on their responses. Potential customers can look for reviews and recommendations to make informed decisions about buying a product or using the service. On the other hand, businesses can use the exercise to take action on relevant feedback from customers in meeting their needs more accurately. Digital marketing is the use of technologies to help marketing activities in order to improve customer knowledge by matching their needs (Chaffey, 2013).

Marketing has been around for a long time. Business owners felt the need to spread the word about their products or services through newspapers and word of mouth. Digital marketing on the other end is becoming popular because it utilizes mass media devices like television, radio and the Internet. The most common digital marketing tool used today is Search Engine Optimization (SEO). Its role is to maximize the way search engines like Google find your website. Digital marketing concept originated from the Internet and search engines ranking of websites. The first search engine was started in 1991 with a network protocol called Gopher for query and search. After the launch of Yahoo in 1994 companies started to maximize their ranking on the website (Smyth 2007). When the Internet bubble burst in 2001, market was dominated by Google and Yahoo for search optimization. Internet search traffic grew in 2006; the rise of search engine optimization grew for major companies like Google (Smyth 2007). In 2007, the usage of mobile devices increased the Internet usage on the move drastically and people all over the world started connecting with each other more conveniently through social media. In the developed world, companies have realized the importance of digital marketing. In order for businesses to be successful they will have to merge online with traditional methods for meeting the needs of customers more precisely (Parsons, Zeisser, Waitman 1996). Introduction of new technologies has creating new business opportunities for marketers to manage their websites and achieve their business objectives (Kiani, 1998). With the availability of so many choices for customers, it is very difficult for marketers to create brands and increase traffic for their products and services. Online advertising is a powerful marketing vehicle for building brands and increasing traffic for companies to achieve success (Song, 2001). Expectations in terms of producing results and measuring 222


success for advertisement money spent, digital marketing is more cost-efficient for measuring ROI on advertisement (Pepelnjak, 2008). Today, monotonous advertising and marketing techniques have given way to digital marketing. In addition, it is so powerful that it can help revive the economy and can create tremendous opportunities for governments to function in a more efficient manner (Munshi, 2012). Firms in Singapore have tested the success of digital marketing tools as being effective and useful for achieving results. (Teo, 2005). More importantly, growth in digital marketing has been due to the rapid advances in technologies and changing market dynamics (Mort, Sullivan, Drennan, Judy, 2002). In order for digital marketing to deliver result for businesses, digital content such as accessibility, navigation and speed are defined as the key characteristics for marketing (Kanttila, 2004). Other tried and tested tool for achieving success through digital marketing is the use of word-of-mouth WOM on social media and for making the site popular (Trusov, 2009). In addition, WOM is linked with creating new members and increasing traffic on the website which in return increases the visibility in terms of marketing. Social media with an extra ordinary example Facebook has opened the door for businesses to communicate with millions of people about products and services and has opened new marketing opportunities in the market. This is possible only if the managers are fully aware of using the communication strategies to engage the customers and enhancing their experience (Mangold, 2009). Marketing professional must truly understand online social marketing campaigns and programs and understand how to do it effectively with performance measurement indicators. As the market dynamics all over the world are changing in relation to the young audience accessibility to social media and usage. It is important that strategic integration approaches are adopted in organization’s marketing communication plan (Rohm & Hanna, 2011). Blogs as a tool for digital marketing have successfully created an impact for increasing sales revenue, especially for products where customers can read reviews and write comments about personal experiences. For businesses, online reviews have worked really well as part of their overall strategic marketing strategy (Zhang, 2013). Online services tools are more influencing than traditional methods of communication ( Helm, Möller, Mauroner, Conrad, 2013). As part of study, it is proven that users experience increase in self-esteem and enjoyment when they adapt to social media which itself is a motivating sign for businesses and marketing professional (Arnott, 2013). Web experiences affect the mental process of consumers and enhance their buying decision online (Cetină, Cristiana, Rădulescu, 2012). This study is very valuable for marketing professional as it highlights the importance of digital marketing. The Internet is the most powerful tool for businesses (Yannopoulos, 2011). Marketing managers who fail to utilize the importance of the Internet in their business marketing strategy will be at disadvantage because the Internet is changing the brand, pricing, distribution and promotion strategy. Pakistan has seen tremendous growth in media with 20 million people have access to the Internet but still marketers insist on doing things the traditional way (Mohsin 2010). Management and structure in Pakistan are still based on ancient paradigm where customers 223


are moving ahead with their demands and expectations. This gap is widening day by day with limited skills and mindset available in Pakistan to solve the problem for the demanding customers. Companies in Pakistan including the MNC’s are going the traditional way and keeping the digital aspect just to show off in tune with the modern trends. 3. METHODOLOGY Sampling: The sample comprising marketing professional in Karachi, Pakistan. Karachi is the biggest city in Pakistan in terms of business presence and commercial activity which is why it was considered for this study. Hundreds of managers were surveyed in Karachi working in different organizations from media, FMCG, Pharmaceuticals, airlines, automobiles, petrochemicals and education. The final sample size was random 200 in which 93% are Men and 17% are Women from the city of Karachi. Research Instrument: This study uses Wilska’s (2003) instrument to measure perceptions of professional. All measures adapted use five-point likert scales. Various non-statistical validity checks were made prior to the questionnaire’s actual implementation. Firstly, all of these constructs were adopted from earlier studies providing acceptably reliable and valid measures. Secondly; these measures had acceptable reliability figures mostly stated in terms of Cronbach’s alpha above 0.5. They have reported a reasonable internal consistency among the items; Cronbach alpha > 0.50 (Wilska, 2003). Finally these measures were processed in a systematic manner in the earlier stages of the research project. In addition to these steps, pre-testing of the questionnaire was also performed. Data Collection: The strategy of using advertising agencies and their clients’ worked really well in terms of questionnaire administration and provided a suitable environment necessary for target participant’s involvement, motivation and convenience. All questionnaires were properly filled and 100% response rate was achieved. Analysis The data was analyzed into ways a) descriptive statistics b) factors analysis. Descriptive Analysis The result from the study indicates that majority of the participants have a perception that digital marketing is a new mix for promotion but also have a negative perception that digital marketing can be misleading and is not useful for word of mouth (WOM) (See Table 1) Perceptions Digital Marketing … is a new avenue for promotion mix. … may provide content not in line with our believes. … can be misleading. … rewrites contents for privacy issues. … accelerates revenue growth. … has low investment. … provides customer's participation. … generates immediate response from customers. … attracts attention very quickly … is much more measurable.

224

M

SD

2.59 2.59 2.51 2.50 2.31 2.31 1.91 1.91 1.86 1.85

.816 .816 .750 .750 .726 .726 .455 .455 .426 .398


… creates marketing opportunities. … useful for word of mouth (WOM). Table 1 Perceptions towards digital marketing

1.85 1.85

.398 .398

Perceptions towards digital marketing tools and their effectives, it was found that mobile phone in terms of SMS and MMS having the highest value followed by online videos, goggle ranking, website content, YouTube and Facebook. All these tools are considered most important for implementing digital marketing practices. Surprisingly, In-depth understanding of technical tools of digital marketing such as Webinars, pay-per-clicks, Google analytics, Blogs and META tags scored low indicating lack of application of these tools and their understanding. (See Table 2) Digital Marketing Tools

M

SD

Mobile Phone – MMS 4.28 .450 Mobile Phone – SMS 4.28 .450 Online Videos 4.28 .450 SEO - Google Rankings 4.28 .450 SEO - Keywords Tags 4.28 .450 Website Contents 4.28 .450 Youtube 4.28 .450 Social Media - Facebook 4.03 .412 Social Media – linkedIn 4.03 .412 Social Media –Twitter 4.03 .412 Webinars 2.84 .943 Pay-per-click 2.83 .941 Google Analytics 2.31 .726 Inlinks 2.31 .726 Blogs 1.85 .398 E-Newsletters 1.85 .398 SEO - Title Tags 1.25 .431 SEO - META Tags / descriptions 1.25 .431 Table 2 Perceptions towards digital marketing tools and their effectiveness

Factor Analysis of Perceptions towards Digital Marketing The data was analyzed in a number of stages. Firstly, exploratory factor analysis was used to determine the factor structure of items related to marketing professional perception towards digital marketing. Secondly, summated score was calculated for resultant digital marketing factors and finally individual differences were measured for marketing professional mindset factors. Factor analysis was conducted for the digital marketing perception mindset scale using a multi-step process which includes three steps; (a) extracting the factors; (b) labeling the factors; c) creating summated scales and examining the descriptive statistics. Analysis of 12 items related to the digital marketing perception scale, using the maximum likelihood method of extraction with direct oblimin rotation, yielded, a four-factor solution, to which various criteria were then applied for refinement. Initially, the solution was examined to determine whether all the factors satisfied the Kaiser criterion (eigenvalues (1) and they did. All the items loading on each separate factor were found to cohere to some 225


degree, and therefore they were included in their respective factors. The above analysis resulted in a final four-factor solution, comprising 12 items, all with communality values greater than 0.3. (See Table 3)

Items M SD Digital marketing …. … is a new avenue for promotion mix. … may provide content not in line with our believes. … rewrites contents for privacy issues. … can be misleading. … is much more measurable. … creates marketing opportunities. … useful for word of mouth (WOM). … provides customer's participation. … generates immediate response from customers. … attracts attention very quickly … has low investment. … accelerates revenue growth.

Skeptic al 2.54 0.74

Factors Enthusia Utilitari st an 1.85 1.89 0.39 0.40

Parsimonio us 2.31 0.72

0.95 0.95 0.94 0.93 0.95 0.95 0.95 0.92 0.92 0.72 0.98 0.98

Factor 1 was labeled as ‘Skeptical’. This group is more skeptical about the importance and benefits of digital marketing. They agree up to certain extent that digital marketing is useful tool for promotional but on the other hand they also think that digital marketing also leads to privacy and misleading of information issues. They have the highest mean value of 2.54 and standard deviation of 0.74. Factor 2 was labeled as ‘Enthusiast’. These professionals have been defined as enthusiast with digital marketing concepts and excited to include them for marketing success. They have a view that digital marketing is useful for creating marketing opportunities and have a positive outlook. They have the lowest mean value of 1.85 and standard deviation of 0.39. Factor 3 was labeled as ‘Utilitarian’. It reflects persons who are most utilitarian in nature and more usage oriented. They use digital marketing services in their routine matters and it is something for them having important in their marketing professional job. The analysis shows that they are the keen user of digital marketing and they are keen on using them as a utility in their professional marketing job. They are more concerned about the utility or usefulness of digital marketing concept and tools. They have the third highest mean value of 1.89 and standard deviation of 0.40. Factor 4 was labeled as ‘Parsimonious’. It reflects a marketing professional who considers that digital marketing is important in terms of cost saving but also gives high importance for growth. They have the second highest mean value of 2.31 and standard deviation of 0.72. 226


Results/Findings The result suggests that professionals in Pakistan are more skeptical towards digital marketing tools and concepts. They do not fully understand the benefits of digital marketing in terms of growth and cost effectiveness. Parsimonious group is more in favor of cost factors of digital marketing and considers it an important tool for growth. This segment of marketing professionals is using the digital marketing strategies and reflects new knowledge and training of professional in Pakistan. 4. CONCLUSIONS This survey examined the perception towards digital marketing of marketing professionals in Pakistan. Although, digital marketing tools and concepts are taking over traditional methods of marketing internationally, it is still a new field for professionals operating in Pakistan. According to this survey, professionals are skeptical about the usage and benefits of digital marketing and have been classified as Skeptical. They do consider it as an important tool for promotion but at the same time concerned about the issues of privacy and misleading of information of digital marketing. SMS and MMS are considered as the most important tool for conducting digital marketing which shows lack of understanding and in-depth usage of digital marketing tools by marketing professionals in Pakistan. 5. REFERENCES AJ Parsons, M Zeisser, R Waitman, “Organizing for digital marketing”, McKinsey Quarterly, 1996 2. Boyd, D. M. & Ellison, N. B. 2007. “Social Network Sites: Definition, History and Scholarship”, Journal of ComputerMediated Communication 13 (1), 210-230. 3. G. Reza Kiani, (1998) "Marketing opportunities in the digital world", Internet Research, Vol. 8 Iss: 2, pp.185 – 194. 4. YS Wang, TI Tang, JE Tang, “An instrument for measuring customer satisfaction toward web sites that market digital products and services”, Journal of Electronic Commerce Research, VOL. 2, NO. 3, 2001 5. A Sundararajan, Leonard N., “Pricing Digital Marketing: Information, Risk Sharing and Performance”, Stern School of Business Working NYU, 2003 6. DC Edelman , “Four ways to get more value from digital marketing”, McKinsey Quarterly, 2010 7. YB Song, “Proof That Online Advertising Works”, Atlas Institute, Seattle, WA, Digital Marketing Insight, 2001. 8. J Chandler Pepelnjak,“Measuring ROI beyond the last ad”, Atlas Institute, Digital Marketing Insight, 2008. 9. A Munshi, MSS MUNSHI, “Digital matketing: A new buzz word”, International Journal of Business Economics & Management Research, Vol.2 Issue 7, July 2012. 10. Thompson S.H. Teo, “Usage and effectiveness of online marketing tools among Businessto-Consumer (B2C) firms in Singapore”, International Journal of Information Management, Volume 25, Issue 3, June Pages 203–213, 2005. 11. Mort, Gillian Sullivan; Drennan, Judy, “Mobile digital technology: Emerging issue for marketing”, The Journal of Database Marketing”, Volume 10, Number 1, 1 September 2002 , pp. 9-23. 1.

227


12. Rick Ferguson, "Word of mouth and viral marketing: taking the temperature of the hottest trends in marketing", Journal of Consumer Marketing, Vol. 25 Iss: 3, pp.179 – 182, 2008. 13. Dickinger, Astrid, “An investigation and conceptual model of SMS marketing”, System Sciences, Proceedings of the 37th Annual Hawaii International Conference, 5-8 Jan, 2004. 14. Nina Koiso-Kanttila, “Digital Content Marketing: A Literature Synthesis”, Journal of Marketing Management, Volume 20, Issue 1-2, pg-45-65, 2004. 15. Michael Trusov, Randolph E. Bucklin, Koen Pauwels (2009). Effects of Word-of-Mouth Versus Traditional Marketing: Findings from an Internet Social Networking Site. Journal of Marketing: Vol. 73, No.5,pp.90-102. 16. Glynn Mangold, David Faulds, “Social media: The new hybrid element of the promotion mix”, Business Horizons, Volume 52, Issue 4, , Pages 357–365, July–August 2009. 17. Hanna, Rohm, Crittenden, “We’re all connected: The power of the social media ecosystem”, Business Horizons, Volume 54, Issue 3, Pages 265–273, May–June 2011. 18. Guoying Zhang, Alan J. Dubinsky, Yong Tan, “Impact of Blogs on Sales Revenue”, International Journal of Virtual Communities and Social Networking, Vol .3, Pg 60-74, Aug-2013. 19. Roland Helm, Michael Möller, Oliver Mauroner, Daniel Conrad, “The effects of a lack of social recognition on online communication behavior”, Computers in Human Behavior Vol 29, pg 1065-1077, 2013. 20. Pai. P, Arnott. DC, “User adoption of social networking sites: Eliciting uses and gratifications through a means–end approach”, Computers in Human Behavior, Volume 29, Issue 3, Pages 1039–1053, May 2013. 21. Cetină. J, Cristiana. M, Rădulescu. V, “Psychological and Social Factors that Influence Online Consumer Behavior”, Procedia - Social and Behavioral Sciences, Vol 62, Page 184188, 2012. 22. Yannopoulos. P, “Impact of the Internet on Marketing Strategy Formulation”, International Journal of Business and Social Science, Vol. 2 No. 18; October 2011. 23. Smyt.G, “The History of Digital Marketing”, Inetasia, 2007. 24. Mohsin. U, “The Rise of Digital Marketing in Pakistan”, Express Tribune, June 21, 2010. 25. Chaffey. D, “Definitions of Emarketing vs Internet vs Digital marketing”, Smart Insight Blog, February 16, 2013.

228


IMAGE REPRESENTATION USING PHOTONS Andreea-Mihaela Pintilie1 Mihai Zaharescu 2 Ion Bucur3 ABSTRACT Genetic brain maps can help physicians to discover patterns of brain structure and how it changes in disease or reacts to medication. In order to generate a brain map, an image model is needed. A much discussed subject, nowadays, is the subject of improving images obtained from devices like MRI, PET, CT etc. In medical area there is a need to improve image segmentation and image resolution; images might be blurred or might contain noise due to the patient’s movement during the process of acquiring them. Imaging studies of the human brain at active medical institutions today routinely accumulate more than 5 terabytes of clinical data per year. The present paper concentrates on neurological field: brain imaging and on genetic field based on the results of the brain imaging. This paper proposes a new medical image format representation, using photons, a tree-like structure in order to improve the inefficiency problem on large medical datasets and an algorithm for eliminating noise from images. KEYWORDS: Image representation, medical image acquisition, image formats, raster images, vector images, photon images, quad trees, binarization, segmentation, edge detection 1. INTRODUCTION Image representation is important when trying to analyze different images. Nowadays medical image is a very important component when trying to establish a diagnostic, when planning an evaluating surgical and radiotherapeutic / chemotherapeutic treatments. There are different registration methods for medical images, based on the purpose of the investigation the medicinal doctor is conducting for a patient. In the following section we provide a short classification of the registration methods in medical imaging area. [1] One of the first criterions is the dimensionality. The main division is based on the spatial time dimensions. We can register two 3D/3D images - obtained from two tomographic data sets with no time dimension or 2D/2D images - obtained as separate slices from tomographic data, or even a 2D/3D registration used for the alignment of spatial data to projective data, where the first image can be a CT one, while the second one a X-ray. Time 1 Engineer, Department of Computer Science and Engineering, Faculty of Automatic Control and Computers Science, University “Politehnica” of Bucharest, Splaiul Independenţei 313, Bucharest, 060042, Romania, [email protected] 2 Engineer, Department of Computer Science and Engineering, Faculty of Automatic Control and Computers Science, University “Politehnica” of Bucharest, Splaiul Independenţei 313, Bucharest, 060042, Romania, [email protected] 3 Associate Professor PhD Eng., Department of Computer Science and Engineering, Faculty of Automatic Control and Computers Science, University “Politehnica” of Bucharest, Splaiul Independenţei 313, Bucharest, 060042, Romania, [email protected]

229


series of images are important when trying to monitor a tumor growth or the effectiveness of a treatment as a pre or post surgical monitoring of healing in different time intervals (short/medium/long). Registration can furthermore be divided into extrinsic registration or intrinsic methods. When acquiring an extrinsic image usually artificial, foreign, invasive objects are used; less invasive objects or even non-invasive markers can be used but with lower accuracy. As an example, before a neurosurgery operation, a stereotactic frame can be used for localization and guidance. Intrinsic methods rely mainly on patient generated image content only. Landmarks for example can be locatable points of the morphology of the visible anatomy or geometrical points providing geometric property such as corners, local curvature extremes etc. Usually, the set of identified points is compared to the original image content and they are used to measure distance. Another intrinsic method is segmentation: same anatomical structures are extracted from two different images and used for the alignment procedure. One image can be elastically deformed in order to be compared with the second image: can be zoomed in or out. The voxel property registration method is different from the other two intrinsic methods described above because the method is applied on the image gray values. Ironically, another registration method is the non-image based one. In this type of registration, two devices are calibrated to each other and that the patient must stay motionless between the first and the second image acquisition. Another classification of the registration task is based on the modalities that are involved. In monomodal registration - the images belong to the same modality, while in multimodal the images are provided from two different modalities. In monomodal method only one image is involved and the other one is the patient himself or a model. As an example, when trying to analyze the myocardial dysfunction in a patient two images can be acquired: under rest and under stress condition. In multimodal method an image obtained from a PET device can be registered to an MR image in order to relate an area of dysfunction to anatomy or the registration of an MR brain image to a defined model of gross brain structures. When a single patient provides the images involved, the method is referred as intrasubject registration, when accomplished using images from distinct patients (or a model and a patient) this is known as intersubject registration; when one image is constructed from a database, an ontology or a knowledge base and the other one is provided by a single patient, the method is called atlas registration. All these image registration methods require different transformations:  rigid: when only translation and rotations are allowed  affine: if the transformation maps the parallel lines onto parallel lines  projective: transformation maps lines on lines  curved: transformation maps lines to curves  global: if the transformation is applied to the entire image  local: if subsection of the image have their transformations defined In general rigid and affine transformations are used as global transformations and

curved transformations are local. Affine transformations are used in rigid body 230


movements, when the scale factors applied on the image are suspected to be incorrect or unknown: as an example, during an MR image acquisition of the lungs: the patient might breath - which will modify the lungs volume and will provide a distorted image. The projective transformation type can be used in 2D/3D image registration as a constrained elastic transformation when a fully elastic transformation behaves inadequately. 2. MOTIVATION In medical image acquisition it is important to have a history in order to establish the rate of progress / regress of a disease. Therefore, it is important to be able to map or to compare different images containing the same anatomical structure. The image file format is defined as the standardized means of organizing and storing digital images. Image files are composed of either pixel or vector (geometric) data that are rasterized to pixels in graphic display (such as monitor, print and other devices). The pixels that constitute an image are ordered in a grid; each pixel has magnitudes of brightness (expressed by gray-scale value) and color. The file size increases directly with the number of pixels composing an image and the color depth of the image: the greater the number of columns and rows, the greater the image resolution will be, hence the size. The size also increases when the pixels color depth increases: an 8 bit pixel will store 256 colors, while a 24-bit pixel will store 16 million colors. The compression method associated with the file format is also important when trying to establish the image size. The images are classified into three families of graphics: raster, vector and metafile (which combine raster and vector information). From the common image file formats we mention JPEG (compression by eliminating nonvisible frequencies, using the cosine transform, and storing color with fewer bits than luminance) GIF (format limited to an 8-bit palette (256 colors); they are suitable for storing graphics with relatively few colors like shapes, diagrams and not for natural images) TIF (saves 8 bits or 16 bits/color for a 24 and 48-bit total) PNG (provides simple detection of common transmission errors as well as full file integrity checking. The format can support up to 16 million colors in a lossless format). 3. RAW (RAW IMAGE FORMAT) The RAW formats use a lossless or nearly lossless compression and produce full size processed images from the same cameras. ISO 12234-2 represents a standard for the raw image format, but most cameras are not standardized or documented. Raw images are also known as digital negatives in film photography. After the acquisition, the image is processed by a raw converter to “positive” file format for further manipulation, which often encodes the image in a device-dependent colorspace. Many medical devices use RAW data format such as Metafile. Metafile represents a generic term for a file format that can store multiple types of data: raster, vector and even type data. The common use is to provide support for the OS computer graphics.

231


DICOM (Digital Imaging and Communications in Medicine) DICOM represents a standard for storing, handling, transmitting and printing medical imaging. Besides being an image format standard, it also represents a network communication protocol. The protocol uses TCP/IP for communication between systems. The DICOM data format contains a structure which allows maintaining not only image information, but as well as other patient information and medical identification. A DICOM data object consists of a number of attributes, including items such as name, ID, etc and also one special attribute containing the image pixel data. A single DICOM object can only contain one attribute containing pixel data. Pixel data can be compressed using a variety of standards, including JPEG, JPEG Lossless, and JPEG 2000. When acquiring the images the time and the spatial dimensions are important. As a consequence algorithms that map a 3D image onto a 2D image must be used. Hence, images must support transformations. We decided to provide a new image format that can support different geometrical transformation and compression and that will maintain its quality when geometrical transformations are applied. 4. OBJECTIVES This paper is mainly concerned with finding an image format that can preserve the image quality event after different geometric transformations are applied on the image. A good example would be a patient who is having multiple RMI, representing different stages in the tumor’s evolution. In order to provide a probability based on the scan images for the malignity of the tumor it is necessary to analyze the images, but analyzing images that have different dimensions, different intensities or that might even contain the object of interest in different positions can be a very difficult task. A map reduce algorithm will be used in order to improve the algorithm computational speed. In order to extract different objects from the image, a canny edge detection algorithm will be used. Another interesting aspect is extracting the tumor from a Pet or RMI image. The main drawback would be sensitivity to false edges, but usually the intensities of the tumor cells are higher than the ones of the normal brain cells and they can be better preserved after considering the luminosity of the pixels around them. 5. IMAGE REPRESENTATIONS Image representation based on spatial relationships In article “Similarity Searching for Chest CT Images Based on Object Features and Spatial Relation Maps”[2] is presented an object base image retrieval system for CT images. The paper proposes an image segmentation method which combines the anatomical knowledge of the chest and the well-known watershed segmentation algorithm. The purpose of segmentation is to identify the mediastinum and the two lung lobes in a chest CT image. The proposed solution has a first step the construction of ARGs (attributed relational graphs) which are used to describe the features of segmented objects.

232


Traditional text-based image retrieval system retrieves images by text keywords. The keywords used cannot reliable describe the features of the images, such as shapes and textures. Several methods have been proposed to approach the CBIR content-based image retrieval problem. To archiving and retrieving images by content, the images need to be: analyzed, indexed and stored to database. Analysis of images can be divided into several steps, like image segmentation, feature extraction[14][15], etc. The contents of a medical image are complex, therefore automatic image segmentation cannot be easily achieved. Features of images can be color, shape and texture of image content [3]. Because most medical images are represented in gray levels, shape and texture features are more often used in describing medical images. Moreover, the spatial relationships among different objects were proposed to further handle multiple objects in an image [4]. In order to achieve a good representation of separated objects in an image, different strategies should be considered for different kinds of images, especially in processing of analyzing medical images acquired from different devices and/or from different parts of the body. The first step is the preprocessing module. The CT image files store the pixel-values as CT numbers. The CT number represents the relative X-ray absorption properties of different organs or tissues with respect to water [5]. Most physiological organisms have positive CT numbers. The air has a negative CT number of -1000. In a typical chest CT image, air occupies much of the space in the lung lobes, hence the property to separate lung lobes (and air) from other parts in the images. The second step is the Modified watershed segmentation algorithm. The method is based on the geographic phenomenon of water flooding up in area with hills and valleys: the water is filling up from the lowest to the highest valleys. The borders between two neighboring valleys are detected when the water in them merge. The watershed segmentation algorithm proposed is the following: 1. The input pixel values are sorted. The purpose is to access the pixels in the same gray level efficiently. 2. Different regions are labeled when the flooding step is processed. 3. The image is segmented in small regions. 4. A region merging step is proceeded to acquire meaningful objects out of enormous small segments. The article proposes 4 merging steps: 1. Merging of the air part with the background: merge regions whose mean pixel values are less than threshold T and are contact with the image background. 2. Merging of regions with similar CT numbers: merge two neighboring regions if the difference in CT number is less than a threshold of 40. 3. Merging of lung lobes: merge regions surrounded by the labels as lung lobes. 4. Merging of mediastinum: merge regions positioned between the sternum, spinal cord, and two lung lobes. After the four merging steps, the three objects (mediastinum, left and right lung lobes) in the chest are identified. After the segmentation step, the objects can be labeled. A spatialrelationship model was used to describe the object features and the relations among different objects. Three properties are used to describe the spatial relationships between two objects: 233


1. D (Distance): the minimum distance between the two objects 2. P (Relative Position): the angle between the connection of the two mass centers and the horizontal line. 3. R (Ratio): the ratio of the areas of the two objects Vector Image representation Vector Image representation is important because of the vector-based graphical content properties: of being editable and scalable. The main trend in vector image representation is developing scalable (resolution-independent) representations of full-color raster images. In a usual path the image plane is decomposed in triangular patches with potentially curved boundaries, and the color signals over the image plane are handled as height fields. The image representation is based on piecewise smooth subdivision surfaces. The image is modeled with discontinuity curves for providing the continuity conditions useful in vectorization ad subsequent vector editing operations. Vectorization techniques:  Triangulation: In representations based on triangulation, each curvilinear feature is approximated by multiple short line segments. This form of representation is not resolution independent because the difference between a smoothly curved feature and a polyline with only C0 continuity at the vertexes become obvious when zoomed in. A technique to overcome the problem is by fitting subdivision curves to patch boundaries that have C2 continuity.  Parametric patches: When grids or Bezier patches are involved, it is necessary to use techniques that are based on higher-order parametric functions. An example of a vectorization technique based on optimized gradient meshes is manually aligning mesh boundaries with salient image features. The main drawback is the fact that mesh placements can be time-consuming. Meshes might introduce color discontinuities that can be approximated using degenerate quads, but the real challenge it to align color discontinuities with the image features. [7]  PDE solutions: The PDE solution is a mesh-free image representation. The technique is based on diffusion curves [6] and it relies on curves with color and blur attributes as boundary conditions of a diffusion process. The color variations of a vector image are represented by the final solution of the diffusion process. The main limitation is that the curves are not coupled together; therefore it is difficult to edit the image using region-based color or shaping operations. Quadtree representation Quad trees are mainly used for image representations because they reduce the storage space of images and the time required for image manipulations. The quad tree structure is efficient to store 2D images. Maintaining a quad tree as an image representation makes it easier to cluster images based on their characteristics such as color, shape, semantics or texture or even by their history. To represent an image by a quadtree the image is recursively split in four disjoint quadrants or squares having the same size. The root node represents the initial quadrant containing the whole image. Different criteria can be used to define the homogeneity property: pixel color, same texture etc. If the image is not homogeneous (according to the criterion established above), the quadtree root has for descendant nodes 234


which represent the four first levels of the image: northwestern, northeastern, southwestern, and southeastern. A node is a leaf when its corresponding image is homogeneous. Generally the quad tree is unbalanced. Each quad tree node has a key, or a location code/quad code. The quad tree node and the corresponding image have the same key. The main operations applied on quad trees are reading and modifying an image quad tree, comparing two quad trees and determine the Q-similarity distance. When trying to modify an image quad tree, one can use the complement operation: each leaf node is accessed and has its value changed. As an example, assuming that the image has a total of 256 different intensity levels to represent a pixel p, the value of p will be changed with (255 –p). In the case of a binary image the black and white values are interchanged. When comparing two quad trees, nodes with the same identifier are actually compared. The basic operations that can be applied on two quad trees are: union, intersection, comparison and difference). When trying to compare a gray-scale or colored image, a similarity measure based on image features is necessary. The node values might be considered as similar if the similarity distance between those two nodes is under a given threshold. The Q similarity distance is defined as the distance between two quad trees. It represents the number of nodes having the same identifier and different values divided by the cardinal of the union of node identifiers. 6. EDGE DETECTION ED represents a terminology in image processing, which refer to algorithms that have the scope to identify points in a digital image at which the image brightness changes sharply or has discontinuities. The next section will present some of the methods used for edge detection. Robert's Cross Operator The Robert’s cross operator performs a simple, quick to compute, 2 – D spatial gradient measurement on an image. It thus highlights regions of high spatial frequency which often corresponds to edges. Pixel values at each point in the output represent the estimated absolute magnitude of the spatial gradient of the input image at that point. In theory the operator consists of a pair of 2×2 convolution kernels as shown below. One kernel is simply the other rotated by 90°.

These kernels are designed to respond maximally to edges running at 45° to the pixel grid, one kernel for each of the two perpendicular orientations. The edge gradient magnitude has the formula: The angle of orientation of the edge giving rise to the spatial gradient is given by:

235


Sobel Operator The Sobel operator performs a 2–D spatial gradient measurement on an image and typically it is used to find the approximate absolute gradient magnitude at each point in an input image. In theory the operator consists of a pair of 3×3 convolution kernel as shown below.

These kernels are designed to respond maximally to edges running vertically and horizontally relative to the pixel grid, one kernel for each of the two perpendicular orientations. The edge gradient magnitude is given by: The angle of orientation of the edge giving rise to the spatial gradient is given by:

Prewitt Compass Edge Detector Compass edge detection is an alternative approach to the differential gradient edge detection. When using compass edge detection the image is convolved with a set of (in general 8) convolution kernels each of which is sensitive to edges in a different orientation. For each pixel the local edge gradient magnitude is estimated with the maximum response of all 8 kernels at this pixel location. Where Gi is the response of the kernel ‘i’ at the particular pixel location and ‘n’ is the number of convolution kernels. The local edge orientation is estimated with the orientation of the kernel that yields the maximum response. Two templates out of the set of 8 for the Prewitt kernel is given below.

The whole set of 8 kernels is produced by taking one of the kernels and rotating its coefficients circularly. Each of the resulting kernels is sensitive to an edge orientation ranging from 0 to 315 insteps of 45, where 0 corresponds to a vertical edge. Zero Crossing Detector The zero crossing detector looks for places in the Laplacian of an image where the value of the Laplacian passes through zero, i.e. points where the Laplacian changes sign. Such points often occur at edges in images, i.e. the points where the intensity of the changes rapidly, but 236


they also occur at places that are not as easy to associate with edges. It is best to think of the zero crossing detector as some sort of feature detector rather than as a edge detector. The core of the zero crossing detector is the Laplacian of Gaussian filter. Once the image has been Laplacian of Gaussian filtered, it only remains to detect the zero crossings. This can be done in several ways. The simplest is to threshold the Laplacian of Gaussian output to zero and a more accurate approach is to perform some kind of interpolation to estimate the position of the zero crossing to sub–pixel precision. Canny Edge Detector The Canny operator was designed to be an optimal edge detector. The Canny operator works in a multistage process. First of all the image is smoothed by Gaussian convolution, then a derivative operator is applied. The resultant edges pass through a first threshold and in order to connect broken edges, pixels that are able to fill the spaces are passed through a second, lower, threshold. Using the edge directions, the edges are thinned to a pixel wide. The effect of the Canny operator is determined by three parameters: the width of the Gaussian kernel, the lower and the upper thresholds used by the tracker. 7. BINARIZATION A binary image is an image that only uses two values for its pixel values. Binary images are often used in digital image processing as masks and are the result of certain operations such as edge detection, segmentation, thresholding. The following section will present different binarization methods [13].  Histogram-based methods: These methods determine the binarization threshold by analyzing the shape properties of the histogram, such as the peaks and valleys. Pavlidis [11] constructs a histogram by using gray-image pixels with significant curvature, or second derivative, and then selects a threshold based on the histogram.  Clustering-based methods: The threshold is selected by partitioning the image’s pixels into two clusters at the level that maximizes the between-class variance, or minimizes the misclassification errors of the corresponding Gaussian density functions [9].  Entropy-based methods: These methods employ entropy information for binarization [8].  Object attribute-based methods: These methods select the threshold based on some attribute quality (e.g., edge matching of Hertz and Schafer [10]) or the similarity measure between the original image and the binarized image.  Spatial binarization methods: These methods binarize an image according to the higher-order probability or the correlation between pixels.  Hybrid methods: Local, adaptive methods [12]; algorithms for improving mentioned methods [13] 8. IMPLEMENTATION The project implementation is in Java using Hadoop Map Reduce for image processing. The first step of the algorithm was to apply the Canny edge detection algorithm on a image. Then we used an entropy-based binarization algorithm as presented in Binarization section to remove the noise. We spread the photons on the edges using an iterative algorithm. The 237


photons were spread based on the fact that the edges will absorb the light, therefore the photons will model the edges. The image processing was done using map reduce to improve the computational speed. We also maintained a quad tree containing the image which provides access with a higher level of detail once the tree is traversed. 9. ACKNOWLEDGEMENT The authors would like to thank Costin-Anton Boiangiu for his original idea (paper under review) onto which the described system is based, for support and assistance with this paper. 10. REFERENCES [1] [2] [3] [4] [5] [6] [7] [8]

[9] [10] [11] [12]

[13]

[14]

[15]

238

J.B.A. Mainz, M.A. Viergever: “A survey of medical image registration”, Conf Proc IEEE Eng Med Biol Soc. 2004;2:1298-301. Yu SN, Chiang CT, “Similarity searching for chest CT images based on object features and spatial relation maps”. Yihong GongIntelligent Image Databases: Towards Advanced Image Retrieval, Robotics Institute, Carnegie Mellon University. Euripides G.M. Petrakis, “Design and evaluation of spatial similarity approaches for image retrieval”, Image and Vision Computing, Volume 20, Issue 1, 1 January 2002, Pp 59–76. Louis M. Castanier, “An Introduction To Computerized X-Ray Tomography For Petroleum Research”, May 1989, Stanford University Petroleum Research Institute. Alexandrina Orzan, Adrien Bousseau, Holger Winnemoller, Pascal Barla, Joelle Thollot, David Salesin, “Diffusion Curves: A Vector Representation for Smooth-Shaded Images”. Tian Xia, Binbin Liao, Yizhou Yu, “Patch-Based Image Vectorization with Automatic Curvilinear Feature Alignment”. J.N. Kapur, P.K. Sahoo, A.K.C. Wong, “A new method for gray-level picture thresholding using the entropy of the histogram”, Computer Vision, Graphics, and Image Processing, Volume 29, Issue 3, March 1985, Pages 273–285. J. Kittler, J. Illingworth, “Minimum error thresholding”, Pattern Recognition, V 19, I 1, 1986, Pp 41–47. L. Hertz, R. W. Schafer, “Multilevel thresholding using edge matching”, Journal Computer Vision, Graphics, and Image Processing archive, Vol44 Is 3, Dec. 1988, Pp 279-295. Pavlidis, T. “Threshold selection using second derivatives of the gray scale image”, Document Analysis and Recognition, 1993. Costin-Anton Boiangiu, Alexandra Olteanu, Alexandru Victor Stefanescu, Daniel Rosner, Alexandru Ionut Egner (2010). „Local Thresholding Image Binarization using VariableWindow Standard Deviation Response” (2010), Annals of DAAAM for 2010, Proceedings of the 21st International DAAAM Symposium, 20-23 11 2010, Zadar, Croatia, pp. 133-134. Costin-Anton Boiangiu, Andrei Iulian Dvornic, Dan Cristian Cananau, „Binarization for Digitization Projects Using Hybrid Foreground-Reconstruction”, Proceedings of the 2009 IEEE 5th International Conference on Intelligent Computer Communication and Processing, Cluj-Napoca, August 27-29, pp.141-144, 2009. Costin-Anton Boiangiu, “The Beta-Shape Algorithm for Polygonal Contour Reconstruction”, CSCS14 – The 14th International Conference on Control System and Computer Science, Bucharest, July 2003. Costin-Anton Boiangiu, Bogdan Raducanu. “3D Mesh Simplification Techniques for Image-Page Clusters Detection”. WSEAS Transactions on Information Science, Applications, Issue 7, Volume 5, pp. 1200 – 1209, July 2008.


DATABASE DYNAMIC MANAGEMENT PLATFORM (DBDMS) IN OPERATIVE SOFTWARE SYSTEMS Virgil Chichernea1 Dragos-Paul Pop2 ABSTRACT This paper discusses the opportunity of using cloud computing and cloud service features to reliably store and manipulate data and databases. It proposes a platform, the Database Dynamic Managemnt Platform (DBDMS), which can be used to effectively handle database versioning of both schema and data. Opportunities and advantages that this systems brings are discussed here and a mathematical model if presented an analyzed for the proposed platform. Keywords: cloud computing, cloud storage, dynamic database, schema versioning, data versioning 1. INTRODUCTION In a globalized society with state of the art informational technologies, Operative Software systems (OSS) that companies use are always under stress to adapt to the dynamics of always more sophisticated demands of precise information delivery. These systems need to adapt as soon as possible to deliver information to all online connected users that need to access Databases (DB) used to store information by the systems. Because of these conditions, Databases used by OSS are always undergoing changes and updates both in the data they are storing but also in their structure and schemas, in order to quickly adapt to changes in requirements, at a low cost in order to provide an optimal costbenefit relationship. These are a few of the basic sources of database structure and content updates that require updating and storing new versions of databases: 

   

the evolution of the systems in question determined by the requirements to integrate it in a globalized world; this evolution imposes an update dynamic for OSS in order to provide adequate responses for attribute dynamics: exact information provided timely, anywhere, anytime, to any number of online users; the explosive evolution of new information technologies, imposes the adoption of some precautionary measures in order for OSS to be transferred on new hardware and on new operating systems; the evolution of Database Management Systems (DBMS) performance and data storing opportunities; legal changes in the business environment and the IT&C environment; the list could go on;

1Proffessor,

Ph.D., Romanian-American University, Bucharest, Romania, [email protected]

2

Teaching Assistant, Romanian-American University, Bucharest, Romania, Ph.D. Student, Academy of Economic Studies, Bucharest, Romania, [email protected]

239


The solutions offered by cloud DBMS are different from other locally found systems, because they do not require physical installation and because they can be accessed with just an Internet connection and a Web browser. Regardless of where it is being accessed from (home, work, automobile, anywhere), the DBMS stores and offers at any time the software, documents, files and data that are needed to keep an OSS functioning. This can be seen as the technical means required to keep pace with the ever growing activity of the system in question. The DBDMS offers access from anywhere and any device connected to the Internet computer, mobile phone, tablet - without interruption, without time loss and at affordable prices. In a cloud environment, the DBDMS platform handles in a permanent and continuous way, the updates of structure and content for operational databases, using an operational versions management system. Access for company users for these versions of the operational databases are managed by this platform, removing down times caused by technical and data security requirements, allowing for an increase in productivity for company personnel. The DBDMS platform is an advanced technical solution, without huge extra initial costs (servers, licenses). The costs for servers and storage space decrease significantly, allowing for a flexible configuration of access for authorized users to data stored in different successive versions (content and dynamic structures) following current business needs. In the cloud costs are predictable, easy to measure and always optimized. The DBDMS platform can be adapted to the specifics of the system in question, maintaining its essence unchanged and this way eliminating the barriers that appear in the natural evolution of structures and data contents that are stored in these databases. System upgrades are offered at a periodic rate in order to keep up with the level of competitiveness and be against it. DBDMS can be integrated in a stable, unitary and perfectly operational platform in any system regardless of company size and domain. Any altering of the requested data or data structure determines the transfer of those changes instantaneously to other operational versions, thus saving time because there is no need to input the data again and data query is assured to be easy no matter the version of database they are stored in. The proposed cloud solutions work in pair with other software applications that are in use and that are to be kept in use. The cloud DBDMS platform offers a safe and secure work environment which allows for safe and secure data storage and keeps unwanted users away and data is not lost in the race of changes, number of users or change of files and documents. Access is reserved only to authorized users. The cloud DBDMS solutions are designed to best address the requirements of each department of the system in question and offer dedicated solutions that cover the entire range of the following interest levels: efficient integrated internal process management (data content and structure), streamlining of the operational flow, productivity increase and optimization of the cost-performance relationship. 240


These are a few of the advantages offered by the DBDMS platform:       

 

Reduced costs – there are no initial acquisition, implementation and equipment maintenance costs, because there is no need to create a new infrastructure; Flexibility – the solution is easy to implement, there is no need to install hardware / software, and it can be used from anywhere and any device. Moreover, at any time new users can be added to the system; Scalability – costs are based on effective number of users and updating their number is done immediately; Mobility – the solution can be accessed at any time and from anywhere, from any equipment as long as there is a connection to the Internet; Modularity – each company activity is managed by a specific application. The solution’s models work either in an integrated manner, using a single database, or independently; Security – an advanced level of security is provided for stored data. This data is protected against industrial espionage, theft or definitive loss of data; Integrability – the solution integrates the whole activity of a company in a unique database. The tracking of all financial and accounting documents within the company is allowed, along with processes undertaken on these documents and contribution from each user that has worked on the database; Business process efficiency – The systems allows for organizational objective fulfillment, providing for real time control and efficiency growth by offering decisional support; Coordination – The system allows for activity planning and work flow control in real time.

2. FACILITIES OFFERED BY CLOUD STORAGE Cloud storage is a network of storage for data and data objects (images, text documents, sound and video files) in virtualized areas hosted by a provider. The provider operates with the data centers (BIGD) that are distributed among many servers and locations and a personnel that requires own data that has been stored in memory locations at a price. Data center operators virtualize available resources and provide, by request, memory locations that users can use by themselves to store data and data objects. The security of files is offered by the provider and by software applications that are being used. The cloud storage system can be accessed through a web service application programming interface (API) or through applications that use this API like cloud desktop storage, cloud storage gateway or web-based content management systems. The new concepts of cloud computing and cloud storage, first envisioned by Joseph Carl Robnett Licklider in the 1960s, work with a large array of terms, from which we underline: storage cloud, private cloud storage, mobile cloud storage, public cloud storage, hybrid cloud storage, personal cloud storage, public cloud, cloud backup, cloud enablement, hybrid cloud, cloud services, private cloud, cloud computing, Amazon Simple Storage Service Amazon S3, etc. [5], [6]. Cloud storage provides a virtual IT structure that can evolve as fast as the system in question, offering a generous environment for developing operational software systems for 241


companies and small and medium enterprises. Small companies can use a section of cloud storage and a specialized software (sync) managed by the provider where they can operate data storage and retrieval queries form any authorized mobile device. [4] Data back-up is made outside of company headquarters, on multiple servers, which allows for a better security model in case of force majeure situations like fires and floods. For example, we can mention some of the best known cloud storage systems like Dropbox and SugarSync or a large number of cloud drive systems like Google Drive, SkyDrive and others. A lot of providers offer free space as a starting point, with ranges from 5 to 25 GB and only charge for extra space or bandwidth. Companies and especially small companies that use these services benefit from significant saving in time and money. These are some of the benefits: 

Cost reduction – the host cloud server optimizes the relationship between: computing speed – storage space – time – running costs and provides for a significant save at company level for these level when thinking about running software systems;  Anytime and anywhere – desktop storage allows for cloud storage access of stored files from any authorized device anytime via a specific software application (sync). A user’s files are stored on multiple servers lowering the risk of technical incidents to a minimum;  Easy collaboration – saving and accessing files in cloud storage is available in a multiuser-multitasking regime, so that all authorized users can access the same stored data at the same time;  Risk reduction – cloud storage provides data security by off-site data backup, reducing the risk of virus infections and other cyber-attacks;  Increase in efficiency – after migrating to cloud storage, small companies won’t have problems regarding computing power, storage space or access to specialized software; Notions, concepts and definitions Any software system stores data in records that are organized in files stored on magnetic drives to allow for quick retrieval. Let there be: {R} – array of records {S} – array of address in the storage space where data is written for these records {C} – array of data requests for records in storage A database DB (R, S, C) for a software system is array R, stored on drive S, in order to satisfy the requests found in C (retrieval of requested data in time). Any database (DB) has a structure of R arrays applied over S (expressed by the structure of files which store data) and a content for that data at the time t+t0. Organizing the DB, O {DB {R, S, C}} is defined as both the content and the files in which data is stored. 242


The main objective of any O {DB{R, S, C}} is the optimization of the relation between storage space S and retrieval time t(ci) under the aspect of total cost. In mathematical terms, this objective can be formulated as follows: With M {O {DB{R, S, C}}} being given, identify a structure O{DB{R, S, C}} that can optimize the following relationship: Min {t(ci), Vci C} at the same time with min Cost (O{BD{R,S,C}}) – under the aspect of total cost. A general solution for this problem is difficult to obtain because of the complex structures and the volume of data related to the arrays R, S, C. In order to achieve this objective, a wide range of techniques for organizing O {DB{R, S, C}} have been developed from simple files to state of the art RDBM systems. The operational software system (OSS) needs to provide exact data in time, anywhere, anytime, to any users connected online and, in this context, the contents of the database, as a support for the OSS, is a updated through the operational flows of the system in question. The dynamics of these updates affect both the current contents of the DB (arrays R and S) and the structure of O {DB{R, S, C}}, influenced by the dynamics of the requests in array C. We define a dynamic database (DDB) the array Oi{BD(R,S,C)}, for i=1,2 … n in which Ok{BD(R,S,C)} represents O{BD(R,S,C)} at the point in time t=tk. We define the platform DBDMS as being the software platform that manages the different versions Oi{BD(R,S,C)} of the OSS by using cloud facilities. 3. THE DDBMS PLATFORM FOR OPERATIVE SOFTWARE SYSTEMS Through the information flows of the system in question (companies, central or local state administrative unite, banks, etc.) the arrays R, S and C are always updated (many times the updates being in real time), both at the content and structure levels (files, file structures, keywords, etc.). In [3] aspects related to Boolean algebra for record arrays are presented and a O{BD(R,S,C)} is proposed as a Boolean algebra of all possible answers and the mechanisms specific to large databases (BIGD). Let us follow the dynamics of updates in a OSS and, implicitly, the dynamics of the Oi{BD(R,S,C)} array. In order to secure the normal functionality of the OSS and to avoid system crashes in case of technical incidents, mechanisms for saving and restoring the state of the OSS have been refined. These are among the most known such mechanisms: 

Backups of different versions of DB and transaction files; in case of technical incidents (hardware or software malfunctions) the DB is to be restored base on these files and updated with the transactions from transaction files from the last DB save; 243




For some certain OSSs (like banking systems), along backups the so called mirroring system is used, i.e. the parallel and real time activity of supplemental DB updating mechanisms;  In all classical OSSs, the altering of DB structure is forbidden. The facilities offered by cloud systems and state of the art software technologies leads to new approaches for the processes of updating, saving and restoring DB (R, S, C). The processes of updating the DB according to request dynamics have driven the development of new platforms that can provide updates to data and structures, but also to save / restore processes in real time to the last version of the O{BD(R,S,C)}. In accordance with the dynamics of the OSS in the stage of facilities offered by cloud and new software technologies (laptop, mobile phone, tablet) we observe the following new notations in the DB: R  R + Ri, S  S + Si C  C + Ci Where Ri, Si, Ci represent changes in structure for the DB, meaning either changes in records in some DB files or the addition or deletion of some files. In this new context, using the previous notations, the aspects of dynamic update of the DB, using the mathematical notations from above, can be expressed as follows: Oi{BD(R + Ri,S +Si,C + Ci)}, for i = 1,2,… n Where 

{Ri} – the new structures of the R array (expressed by either changes in the structures of existing DB files or by the addition or deletion of files);  {Si} – the array of addresses of the storage space in which these records are stored;  {Ci} – new requests of the system in question With the new notation let’s consider a number of keywords (fields contained in the DB records), denoted as k1, k2, k3 … kn with the property that any record in {R + Ri} contains at least one ki keyword. We denote with: R(ki) +Ri(ki)} set of records that contain the keyword ki; {A(ki) +Ai(ki)} list of addresses that hold the records found in the array {R(ki) +Ri(ki)}; From [3] we can prove that any request from C for the DB can be written as a Boolean function like: F(k1, k2, k3, …, kn) = Ki The answer to this request for data is found in the record collection B B(R + Ri), where we name the array B(R + Ri) the array of all possible answers.

244


4. TECHNIQUES FOR SAVING / RESTORING OF THE DB IN OSS In DBs of OSSs various files store numerical data and object data (images, text documents, multimedia files) as well as addresses for records that allow the retrieval of this data in time.

The file for direct transactions is an intermediary file that contains all data used to undertake the following operations on the DB:  Inserting new records in the DB certain files;  Deleting records from certain DB files;  Changing records in certain DB files. The file for reflected transaction is a file attached to the direct transaction file of the system in the following way: 

For any record in the direct transaction file there is a record in the reflected transaction file defined as follows: Let us consider: o o o

records of the direct transactions file: 𝑥𝑖 = {𝑑𝑘𝑖 (𝑡), 𝑘 = 1 … 𝑛}; 𝑖 = 1 … 𝑝; 𝑡 ∈ 𝑇 records of the database 𝑥̃𝑖 = {𝑎𝑘𝑖 (𝑡), 𝑘 = 1 … 𝑛}; 𝑖 = 1 … 𝑝; 𝑡 ∈ 𝑇 records of the reflected file 𝑥̅𝑖 = {𝑑̅𝑘𝑖 (𝑡), 𝑘 = 1 … 𝑛}; 𝑖 = 1 … 𝑝; 𝑡 ∈ 𝑇 𝑑̅𝑘 (𝑡) is defined as: 𝑖

̅ 𝑘 (𝑡) = −𝑑𝑘 (𝑡), 𝑖𝑓 𝑎𝑘 (𝑡): = 𝑎𝑘 (𝑡) + 𝑑𝑘 (𝑡) 1. 𝑑 𝑖 𝑖 𝑖 𝑖 𝑖 ̅ 𝑘 (𝑡) = 𝑑𝑘 (𝑡), 𝑖𝑓 𝑎𝑘 (𝑡): = 𝑎𝑘 (𝑡) − 𝑑𝑘 (𝑡) 2. 𝑑 𝑖 𝑖 𝑖 𝑖 𝑖 3. 𝑥 ̅ 𝑖 ∶= 𝑥̅ 𝑖 ∪ {𝑑̅ 𝑘𝑖 (𝑡), 𝑘 = 1 … 𝑛} 𝑖𝑓 𝑥̃ 𝑖 ∶= 𝑥̃ 𝑖 ∩ {𝑑̅ 𝑘𝑖 (𝑡), 𝑘 = 1 … 𝑛} 4. 𝑥 ̅ 𝑖 ∶= 𝑥̅ 𝑖 ∩ {𝑑̅ 𝑘𝑖 (𝑡), 𝑘 = 1 … 𝑛} 𝑖𝑓 𝑥̃ 𝑖 ∶= 𝑥̃ 𝑖 ∪ {𝑑̅ 𝑘𝑖 (𝑡), 𝑘 = 1 … 𝑛} 

Both files will be stored in cloud storage but on different servers and will serve for both monitoring the correct evolution of the DB (by comparing records from transaction T at the moment t1) and for rebuilding the DB in case of a technical incident, by using the backup of the DB at the moment t and undertaking direct and reflected transactions on that backup from the interval t + t1 and comparing the comparing the contents of these two rebuilt DBs by the use of the two rebuild procedures. This way, by using the 3 entities (DB, direct transactions file, reflected transactions file) and by using the procedures for rebuilding the contents of the DB at the moment t + t1 procedures and methods are provided for rebuilding the DB at certain moments its evolution. This is requested on one hand by the possibility of the destruction of the DB (technical incidents, cyber-attacks, etc.) and on the other hand by the frequent reports demanded on the state of the system at a given point in time. 245


Figure 1 DB update and reflected transaction file creation

Figure 2 Restoring de DB at a certain stage

5. CONCLUSIONS Cloud storage facilities for securely storing the three entities allows for designing and maintaining dynamic DBs in service. These DBs can support, besides the classical operations of updating (adding, deleting, changing records), the update of the DB structure, the change of the structure of records in DB files and the addition and deletion of DB files dynamically according to the evolution of the system in question. 6. ACKNOWLEDGEMENT This work was co-financed from the European Social Fund through Sectorial Operational Program Human Resources Development 2007-2013, project POSDRU/107/1.5/S/77213 „Ph.D. for a career in interdisciplinary economic research at the European standards” 7. BIBLIOGRAPHY [1] Cloud storage – Wikipedia the free encyclopedia, www.en. Wikipedia.org/wiki/cloud storage [2] Cloud storage & Unlimited, www. Online File storage –Raskpage Cloud [3] Chichernea, V., Pop, D., “Techniques For Optimizing The Relationship Between Data Storage Space And Data Retrieval Time For Large Databases”, Journal of Information Systems and Operations Management, vol. 6, no.2, pp. 258-268, 2012 [4] Armbrust, M., Fox, A., Griffith, R., Joseph, A., & RH. (2009). Above the clouds: A Berkeley view of cloud computing. University of California, Berkeley, Tech. Rep. UCB , 07–013. [5] Lee, G., Rabkin, A., Katz, R., Stoica, I., Griffith, R., Joseph, A. D., … Zaharia, M. (2010). A view of cloud computing. Communications of the ACM. [6] Mirashe, S. P., & Kalyankar, N. V. (2010). Cloud Computing. (N. Antonopoulos & L. Gillam, Eds.)Communications of the ACM, 51(7)

246


TEXT LINE SEGMENTATION IN HANDWRITTEN DOCUMENTS BASED ON DYNAMIC WEIGHTS Costin-Anton Boiangiu1 Mihai Cristian Tanase2 Radu Ioanitescu3 ABSTRACT Identification of text lines in documents, or text line segmentation, represents the first step in the process called ‘Text recognition”, whose purpose is to extract the text and put it in a more understandable format. The paper proposes a seam carving algorithm as an approach to find the text lines. This algorithm uses a new method that allocates dynamic weights for every processed pixel in the original image. With this addition, the resulting lines follow the text more accurately. The downside of this technique is the computational time overhead. KEYWORDS: OCR, text line segmentation, handwritten documents, dynamic weights 1. INTRODUCTION The process of extracting lines from a document is used as a basis for document structure extraction, handwriting recognition or text enhancement. There are numerous methods ([13], [9], [5]) that address the printed document line extraction problem which is usually reduced to global skew search (the text lines are parallel with each other, but not necessarily horizontal). On the other hand, when dealing with handwritten document the problem becomes more complex ([14], [12], [11], [10], [8], [7], [6], [4]): the lines are not parallel with each other, same letters do not have same sizes, text lines have letters that extend to other text lines, higher text organization cannot be defined (paragraphs, subsections etc.). In any image, document or not, with handwritten or printed text, each pixel can be associated with an importance (i.e., how much that pixel influences the overall image). In this paper the importance of the pixels in an image is given by the energy map, which is further used to compute the energy cost map and, finally, the seam carving algorithm is used to detect the text lines. 2. RELATED WORK There are many techniques that address the problem of text line segmentation. They are generally divided into two categories:  

techniques that work directly on the gray scale image techniques that use, as input, the binary representation of the image.

1 Associate Professor PhD Eng, ”Politehnica” University of Bucharest Romania, 060042 Bucharest, [email protected]. 2 Engineer, VirtualMetrix Design Romania, 060104 Bucharest, ,[email protected] 3 Engineer, European Pattent Office (EPO) Germany, Bayerstrasse 34, Munich, [email protected]

247


For the gray scale images, there are algorithms that use the, so called, projection profiles representing the sum of all the pixels values in a given direction. This method, applied in a pre-computed direction, which usually uses the Hough transform, can accurately identify text lines in printed documents, but fails to produce even moderate results when applied to handwritten documents. Different workarounds of this problem were proposed: statistical division of the text page, which tries to closely follow the local skew of the handwritten text; smearing methods, which fills the space in-between text characters; Hough transform which selects all lines that have the value accumulated in Hough space greater than a given threshold; repulsiveattractive method in which the current line is attacked by the pixels from the text and is repelled by previous found text lines. For the binary images, there are various grouping methods that can be used in addition to the above mentioned techniques. In this paper, the energy map of a gray-scale image is used to compute the energy cost map which, in turn, represents the input for the seam carving algorithm. The result is a combination between the original document and lines that follow the text which represents the segmentation boundaries. Image as an energy map Removing insignificant pixels

Removing information pixels

Before

Before

After Figure 1 Exemplification of the energy map concept

After

Original image

The energy map of an image represents the information quantity map. Each pixel in the energy map has a value associated with it that represents the amount of information that the given pixel stores in the image. If a high energy pixel is removed from the image, the resulting image has a significant drop in detail, whereas removing a low energy pixel results in a negligible information loss. Figure 1 illustrates this concept. Removing a set of pixels, each belonging to a homogeneous area will result in almost insignificant information loss compared to extracting the second set of pixels. From this example, an observation can be made: pixels belonging to homogeneous areas have low energy and pixels belonging to areas with high variations have high energy. This observation leads to the idea of viewing energy map as the derivate of the original image. In [4], the energy map is computed as the distance transform of the binary image (the objects of interest in the image are the text characters). The output of distance transform represents 248


a map of pixels, each assign with the distance from that pixel to the closest text pixel. For this energy map, high energy represents areas furthest from the text and small energy represents text or areas very close to text. Three methods are used for calculating the energy map:  f   f  f        x   y  1. Magnitude of the gradient, namely: 2

2

whereas pixels at the edge receive higher energies. 2. Gaussian 1st derivate 3. Inverse distance transform’ To use distance transform, first, the original image needs to be transformed in a binary image (a binary threshold is applied), than the distance transform algorithm is applied which outputs an image where small values are associated with the text and high values with the background. To have a consistent implementation (i.e., high energy represents high variations), the last image needs to be inverted, which results in an image where the high energy values are associated with text. 3. DYNAMIC WEIGHTS The next step in the process is to calculate the energy cost map. Similar to [11] and [4], this cost map assigns to each pixel a minimum value calculated with the formula below:

M  i, j  2*e  i,j +min  w  neighbors_number/2  k  *M  i  direction, j  k   Where -neighbors_number/2