Texture based Scene Categorization using

ICGST-GVIP, ISSN 1687-398X, Volume (8), Issue (IV), December 2008

Texture based Scene Categorization using Artificial Neural Networks and Support Vector Machines: A Comparative Study Devendran V1, Hemalatha Thiagarajan2, Amitabh Wahi3 Department of Computer Applications, Bannari Amman Institute of Technology, Sathyamangalam, TamilNadu, India 2 Department of Mathematics, National Institute of Technology, Trichy, TamilNadu, India 3 Department of Information Technology, Bannari Amman Institute of Technology, TamilNadu, India Email: [email protected], [email protected], [email protected]

1

under unsupervised circumstances. In [11][12], Bosch et al. present a scene description and segmentation system capable of recognizing natural objects (e.g., sky, trees, grass) under different outdoor conditions. Texture is a commonly used feature in the analysis and interpretation of images. Lance M. Kaplan [13] worked on mosaic images using extended self similar model (ESS) and k-means classification algorithm. Arivazhagan et al., [14][15] worked on classification of mosaic images using statistical features from Ridgelet and Curvelet Transformed Images. G. Y. Chen and P. Bhattacharya [16] worked on invariant texture classification using Ridgelet packets, Fourier transformation and db4 with nearest neighbor classifier. Therefore, the research presented in this article focuses on texture analysis for scene categorization using Artificial Neural Networks (ANN) and Support Vector Machines (SVM). I am in the process of selecting the better feature extraction method which gives the maximum performance with less execution time in the natural scene categorization problems. In my previous work [17], five levels of wavelet decomposition is applied for classification of images using neural classifier and support vector machines. This proves the efficiency of radial basis kernel function in terms of consist performance. In the previous work [18][19], Invariant Moment Features are tested in the scene classification with Neural classifier and Support Vector Machines. In our work [20], haar wavelet, invariant moments and co-occurrence matrices features are tested for their consistency over various binary class problems in natural scenes categories. This shows that co-occurrence features are giving average classification rate of 73.5 % over six various binary classification problems. Hence, this paper presents the natural scene classification using co-occurrence matrix features with polynomial and radial basis kernel function using support vector machines. The results are also compared with the traditional neural networks with backpropagation algorithm. The organization of the paper is as follows: Section 2 describes Texture Features, Section 3 describes Artificial Neural Networks, Section 4 deals with Support Vector

Abstract Categorization of scenes is a fundamental process of human vision that allows us to efficiently and rapidly analyze our surroundings. Scene classification, the classification of images into semantic categories (e.g., coast, mountains, highways and streets) is a challenging and important problem nowadays. This paper uses gray level cooccurrence matrix method to extract features from the scenes and trying to recognize the scene categories called ‘MIT-street’ and ‘MIThighways’. Artificial neural networks and support vector machines classifiers are used for the classification. The comparative results are proving efficiency of support vector machines towards scene categorization problems. The sample images are taken from the real world dataset. Keywords: Artificial Neural Networks, Gray Level Cooccurrence Matrix, Scene Categorization, Support Vector Machine.

1. Introduction In the machine learning literature [1], the term “natural scene” is usually intended as the one of a semantically coherent, namable human-scaled view of an outdoor real world environment, and the term natural scene categorization refers to the task of grouping images into semantically meaningful categories. Understanding the robustness and rapidness of human scene categorization has been a focus of investigation in the cognitive sciences over the last decades [2][3][4]. At the same time, progress in the area of image understanding has prompted computer vision researchers to design computational systems that are capable of automatic scene categorization. Classification is one of several primary categories of machine learning problems [5]. Papers [6], [7] and [8] give very promising results in the classification of indoor-outdoor scene image and manmade-natural classification. Ian Stefan Martin [9] presents in his doctoral work, the techniques for robust learning and segmentation in scene understanding. In [10], Manuele Bicego et al. give a new approach to scene analysis 45

ICGST-GVIP, ISSN 1687-398X, Volume (8), Issue (IV), December 2008

Machine, Section 5 explains the proposed work, Section 6 deals with Implementation of ANN, Section 7 deals with Implementation of SVM, Section 8 deals with Discussion, and finally Section 9 concludes with conclusion.

training samples. This neural network structure used in our experiment is consist of only three layers having 32 neurons in the input layer, 12 neurons in the hidden layer and 1 neuron in the output layer as shown in Figure 2.

2. Texture Features

Table 1 Computation of Texture Features

Texture is a commonly used feature in the analysis and interpretation of images. Texture is characterized by a set of local statistical properties of pixel intensities. We base our texture feature extraction on the spatial graylevel cooccurrence-matrix (SGLCM). The GLCM method considers the spatial relationship between pixels of different gray levels.

Feature

Formula G −1

Energy

∑ ( P θ (i, j ))

i , j =0 G −1

Inertia

The method calculates a GLCM by calculating how often a pixel with a certain intensity i occurs in relation with another pixel j at a certain distance d and orientation θ . For instance, if the value of a pixel is 1 the method looks, for instance, the number of times this pixel has 2 in the right side. Each element (i, j) in the GLCM is the sum of the number of times that the pixel with value i occurred in the specified relationship to a pixel with value j in the raw image.

2

d,

∑ (i − j )

i , j =0

2

Pd ,θ (i, j )

G −1

Entropy

∑ P θ (i, j ) log

i , j =0

G −1

Homogeneity

d,

Once the GLCM is calculated several second-order texture statistics can be computed as illustrated in Table 1 where Pd ,θ (i, j ) is the GLCM between i and j.

Contrast

Cooccurrence matrices are calculated for four directions: 00, 450, 900 and 1350 degrees. The eight Haralick texture descriptors are extracted from each cooccurrence matrices which are computed in each of four angles. Thus 32 features are used to represent an image.

2

max i , j Pd ,θ (i, j ) G −1

G −1

i , j =0

i , j =0

∑ (i − j ) 2 ∑ Pd ,θ (i, j ) G −1

Inverse

[ Pd ,θ (i, j )]

Pd , (i, j )

∑ 1 + (θi − j )

i , j =0

Maximum Probability

2

∑

i , j =0

P d ,θ (i, j ) (i − j ) 2

G −1

Correlation

∑ ijP θ (i, j ) − μ

i , j =0

d,

x

μy

σ xσ y

3. Artificial Neural Networks The Neural networks [21] developed from the theories of how the human brain works. Many modern scientists believe the human brain is a large collection of interconnected neurons. These neurons are connected to both sensory and motor nerves. Scientists believe, that neurons in the brain fire by emitting an electrical impulse across the synapse to other neurons, which then fire or don't depending on certain conditions. Structure of a neuron is given in Figure 1.

Output Layer (1 neuron)

Input Layer (32 neurons)

Hidden Layer (12 neurons)

Figure 2 Simple Neural network Structure

A learning problem with binary outputs (yes / no or 1 / 0) is referred to as binary classification problem whose output layer has only one neuron. A learning problem with finite number of outputs is referred to multi-class classification problem whose output layer has more than one neuron. The examples of input data set (or sets) are referred to as the training data. The algorithm which takes the training data as input and gives the output by selecting best one among hypothetical planes from hypothetical space is referred to as the learning algorithm. The approach of using examples to synthesize programs is known as the learning methodology. When the input data set is represented by

Figure1 Structure and functioning of a single neuron

The Artificial neural network is basically having three layers namely input layer, hidden layer and output layer. There will be one or more hidden layers depending upon the number of dimensions of the 46

ICGST-GVIP, ISSN 1687-398X, Volume (8), Issue (IV), December 2008

its class membership, it is called supervised learning and when the data is not represented by class membership, the learning is known as unsupervised learning. There are two different styles of training .ie, Incremental Training and Batch training. In incremental training the weights and biases of the network are updated each time an input is presented to the network. In batch training the weights and biases are only updated after all of the inputs are presented. In this experimental work; Back propagation algorithm is applied for learning the samples, Tan-sigmoid and logsigmoid functions are applied in hidden layer and output layer respectively, Gradient descent is used for adjusting the weights as training methodology.

separates the data without error, but also maximizes the margin, i.e., maximizes the distance between the closest vectors in both classes to the hyperplane [22] . As shown in Fig. 3, the margin, ρ , is the sum of the absolute distance between the hyperplane and the closest data points in each class. It is given by: ρ = min

| W .X i + b | | W .X i + b | 2 + min = || W || || W || || W ||

(5)

4. Support Vector Machines Support vector machine is a relatively new pattern classifier introduced by Vapnik [22]. A SVM classifies an input vector into one of two classes, with a decision boundary developed based on the concept of structural risk minimization (of classification error) using the statistical learning theory. The SVM learning algorithm directly seeks a separating hyperplane that is optimal by being a maximal margin classifier with respect to training data. For non-linearly separable data, the SVM uses kernel method to transform the original input space, where the data is non-linearly separable, into a higher dimensional feature space where an optimal linear separating hyperplane is constructed. On the basis of its learning approach, the SVM is believed to have good classification rate for high-dimensional data. Consider the problem of image classification where X is an input vector with ‘n’ dimensions. The SVM performs the following operation involving a vector W = ( w1 ,..., wn ) and scalar b:

Figure 3 Optimal separating hyperplane for 2-Dimensional two-class problem

Here, the first min is over Xi of one class and the second min is over Xi of the other class. Therefore, the optimal separating hyperplane is the one that maximizes 2/||W||, subject to constraints (4). It is mathematically more convenient to replace maximization of 2 / ||W|| with the equivalent minimization of ||W||2/ 2 subject to constraints (4), which can be solved by the Lagrangian formulation: l min L = 1 || W || 2 −∑ α i [ y i (W . X i + b) − 1] 2 i =1

f ( X ) = sgn(W • X + b)

where

(1) Positive sign of f(X) may be taken as ‘MIT-street’ images and negative value of f(X) may be regarded as ‘MIT-highways’ images. Consider a set of training data with l data points from two classes. Each data is denoted by (Xi, yi), where i=1, 2,…, l , Xi = ( xi1,…, xin), and yi ∈ {+1, -1}. Note that yi is a binary value representing the two classes. The task of SVM learning algorithm is to find an optimal hyperplane (defined by W and b) that separates the two classes of data. The hyperplane is defined by the equation: (2) W • X +b = 0 Where X is the input vector, W is the vector perpendicular to the hyperplane, and b is a constant. The graphical representation for a simple case of twodimensional input (n = 2) is illustrated in Fig. 3. According to this hyperplane, all the training data must satisfy the following constraints: W • X i + b ≥ +1 for ∀ i = +1 (3) W • X i + b ≥ −1 for ∀ i = -1

is the Lagrange multiplier ( α i >= 0,

i=1,2,.., l ). The Lagrangian has to be minimized with respect to W and b, and maximized with respect to α i . The minimum of the Lagrangian with respect to W and b is given by:

∂L l = 0 ⇒ W = ∑i =1α i X i y i ∂W ∂L l = 0 ⇒ ∑i =1α i y i = 0 ∂b

(7)

(8)

Substituting (7) and (8) into (6), the primal minimization problem is transformed into its dual optimization problem of maximizing the dual Lagrangian L D with respect to α i : Max LD=

which is equivalent to :

yi (W • X i + b) ≥ 1 ∀ i = 1, 2, …., l

αi

(6)

∑

subject to

(4)

l i =1

αi −

∑

l i =1

There are many possible hyperplanes that separate the training data into two classes. However, the optimal separating hyperplane is the unique one that not only

αi ≥ 0 ∀ 47

1 l l ∑ ∑ α iα j yi y j ( X i • X j ) 2 i =1 j = 1

α i yi = 0

i=1,…, l

(9) (10) (11)

ICGST-GVIP, ISSN 1687-398X, Volume (8), Issue (IV), December 2008

Radial basis function:

Thus, the optimal separating hyperplane is constructed by solving the above quadric programming problem defined by (9)-(11). In this solution, those points have non-zero Lagrangian multipliers ( α i > 0) are termed

K(X i , X j ) = e K (X i, X j ) =

i

i

i

• X ) + b)

(13)

Note that, in (13), one only needs to make use of Xi, yi and α i of the support vectors, while X is the input

n

the Xi • Xj term in (9). To avoid the expensive computations of φ ( X i ) • φ ( X j ) in the feature space, it is simpler to employ a kernel function such that (14)

Zero-Mean Normalization: By using this type of normalization, the mean of the transformed set of data points is reduced to zero. For this, the mean and standard deviation of the initial set of data values are required. The transformation formula is v' = (v meanA) / std_devA where meanA and std_devA are the mean and standard deviation of the initial data values. The following are the other techniques normally used by our research community for normalization of data.

(b) Fig. 4 Representation of (a) linearly separable (b) nonlinearly separable Thus, only the kernel function is used in the training algorithm, and one does not need to know the explicit form of φ . The computation in (15) results in some restrictions on the form and parameter values of nonlinear functions that can be used as the kernel functions. Detailed discussions can be found in [22] and [23]. Some commonly used kernel functions are: Polynomial function: d

(18)

In classification, a system is trained to recognize a type of example or differentiate between examples that fall in separate categories. In the case of computer vision, the examples are representations of photographic images and the task of the classifier is to indicate whether or not a specific object or phenomena of interest is present in the image. In order to successfully accomplish this, the classifier must have sufficient prior knowledge about the appearance of the object. This paper is trying to recognize the scenes of two different categories called ‘MIT-street’ and ‘MIThighways’. The complete flow chart of the computerbased system is shown in figure 5. The sample images are taken from the Ponce Research Group [24] which contains 15 different scene categories with 250 samples each. Sample scenes are given in Fig. 7, 8, 9, 10 and 11. Cooccurrence Matrices are used for extracting the features from the images/scenes. Normalization is then applied using Zero-mean normalization method in order to maintain the data within the specified range and also found suitable to improve the performance of the classifier.

function φ : R → H . In H, an optimal separating hyperplane is then constructed using training data in the form of dot products φ ( X i ) • φ ( X j ) instead of

K ( X i , X j ) = (X i . X j + 1)

(17)

5. Proposed Work

vector to be classified. When a linear boundary is inappropriate (i.e., no hyperplane exists to separate the two classes of data), the extension of above method to a more complex decision boundary is accomplished by mapping the input vectors X ∈ Rn into a higher dimensional feature space H through a non-linear

K ( X i .X j ) = φ ( X i ) • φ ( X j )

(16)

The hyperplane and support vectors used to separate the linearly separable data are shown in Fig. 4 (a). And the hyperplane and support vectors used to separate the non-linearly separable data are shown in Fig. 4 (b). Radial basis kernel function with p1=5 used for this non-linear classification. Individual colors represents particular each class of data.

value of b be used in the classification. With this solution, the SVM classifier becomes

∑ y α (X

1+ e

1 [v ( X i . X j ) − δ ]

⎛ ⎞ f ( X ) = sgn ⎜⎜ ∑ y iα i K ( X i , X ) + b ⎟⎟ ⎝ ∀i ,α i >0 ⎠

]

∀ i ,α i > 0

2σ 2

Where d is a positive integer, and, σ , ν and δ are real constants. These four parameters must be defined by the user prior to SVM training. With the use of a kernel function, the SVM capable of performing nonlinear classification of input X becomes,

from (7). b can be obtained by using the Karush-KuhnTucker(KKT) complementary condition for the primal Lagrangian optimization problem: α i y i (W . X i + b) − 1 = 0 ∀ i = 1,.., l (12) One b value may be obtained for every support vector (with α i > 0). Burges[23] recommends that the average

f ( X ) = sgn( W • X + b ) = sgn(

|| xi − x j ||2

Sigmoid function

support vectors. Support vectors satisfy the equality in the constraint (4) and lie closest to the decision boundary (they are circles in Fig. 3, lying on the dotted lines on either side of the separating hyperplane). Consequently, the optimal hyperplane is only determined by the support vectors in the training data. Based on the α i values obtained, W can be calculated

[

−

Decimal Scaling: This type of scaling transforms the data into a range between [-1,1]. The transformation formula is v'(i) = v(i)/10k for the smallest k such that max ( |v'(i)| ) ≤ 1.

(15) 48

ICGST-GVIP, ISSN 1687-398X, Volume (8), Issue (IV), December 2008

Min-Max Normalization: This type of normalization transforms the data into a desired range, usually [0,1]. The transformation formula is v’(i) = ( v(i) – mina ) / (maxA-minA) * (new_maxA- new_minA) + new_minA where, [minA, maxA] is the initial range and [new_minA, new_maxA] is the new range. The normalized data is given to Artificial Neural Networks and Support Vector Machine; and are trained and tested to recognize the scene categories.

network is 32-12-1. Fig. 6 depicts the converging training graph of neural classifier.

Feature Extraction

Input Image

Data Normalization

Classification using ANN

6. Implementation Using ANN Using the above feature vector representation, neural classifier is trained and tested to categorize the scenes. In Training phase, 200 samples are used including 100 samples from ‘MIT-street’ and 100 samples from ‘MIT-highways’. In testing phase, 200 more samples are used including 100 samples from ‘MIT-street’ and 100 samples from ‘MIT-highways’. The input images are of (256x256) pixel size. Zero-mean normalization method is applied to the extracted co-occurrence features. Normalized features are given as input to Artificial Neural Networks to recognize the scene category. Backpropagation algorithm is used to train the neural classifier. The structure of the neural

Classification using SVM

Results Figure 5 Detailed description of proposed work

Fig. 6 Neural Training to classify ‘MIT-street’ and ‘MIT-highways’ scene categories

Fig. 7 Sample images of ‘MIT-highway’ category

Fig. 8 Sample images of ‘MIT-street’’ category

49

ICGST-GVIP, ISSN 1687-398X, Volume (8), Issue (IV), December 2008

Fig. 9 Sample images of ‘industrial’ category

Fig. 10 Sample images of ‘bedroom’ category

Fig. 11 Sample images of ‘mountain’ category

whose results as follows: True Positive (Correctly classified positive sample) is 98 out of 100 and False Positive (Correctly classified negative sample) is 83 out of 100. Status : OPTIMAL_SOLUTION |w0|^2 : 1.294163 Margin : 1.758067 Support Vectors : 23 (11.5%)

7. Implementation Using SVM Kernel function is trained to find the optimal hyperplane to separate two different categories of scenes and maximize the margin between the two classes of data. Polynomial kernel function with degree 2, 3, 4, 5 and Radial basis kernel function with p1=5 are used in SVM for scene classification and the results are compared. In Training phase, 200 samples are used including 100 samples from ‘MIT-street’ and 100 samples from ‘MIT-highways’. In testing phase, 200 more samples are used including 100 samples from ‘MIT-street’ and 100 samples from ‘MIT-highways’. The input images are of (256x256) pixel size. Zeromean normalization method is applied to the extracted features. Normalized features are given as input to Support Vector Machine for training to recognize the scene category.

Polynomial Kernel Function (degree 4): This produces 91.5% classification rate in 184.78 seconds whose results as follows: True Positive (Correctly classified positive sample) is 98 out of 100 and False Positive (Correctly classified negative sample) is 85 out of 100. Status |w0|^2 Margin Support Vectors

Polynomial Kernel Function (degree 2): This produces 90.0% classification rate in 187.75 seconds whose results as follows: True Positive (Correctly classified positive sample) is 95 out of 100 and False Positive (Correctly classified negative sample) is 85 out of 100. Status : OPTIMAL_SOLUTION |w0|^2 : 171.182218 Margin : 0.152862 Support Vectors : 23 (11.5%)

: : : :

OPTIMAL_SOLUTION 0.013841 16.999672 23 (11.5%)

Polynomial Kernel Function (degree 5): This produces 93.0% classification rate in 184.52 seconds whose results as follows: True Positive (Correctly classified positive sample) is 98 out of 100 and False Positive (Correctly classified negative sample) is 88 out of 100. Status : OPTIMAL_SOLUTION |w0|^2 : 0.000243 Margin : 128.277694 Support Vectors : 7 (3.5%)

Polynomial Kernel Function (degree 3): This produces 90.5% classification rate in 187.44 seconds 50

ICGST-GVIP, ISSN 1687-398X, Volume (8), Issue (IV), December 2008

Radial Basis Kernel Function (degree 5): This produces 96.0% classification rate in 197.03 seconds whose results as follows: True Positive (Correctly classified positive sample) is 97 out of 100 and False Positive (Correctly classified negative sample) is 95 out of 100. The optimal hyperplane is given in Fig. 12. Status : OPTIMAL_SOLUTION |w0|^2 : 76084.150160 Margin : 0.007251 Support Vectors : 22 (11.0%) The pictorial representation which shows the comparative study of the performances of artificial neural networks, polynomial and radial basis kernel function are shown in Fig. 13 in terms of classification rate and execution time.

Fig. 12 Optimal Hyperplane that separates ‘MIT-street’ and ‘MIT-highways’ scene categories

900 800

770.47

700 600 500

Classification %

400

Execution Time

300 187.75

200 100

92.5

90

187.44 90.5

184.78

197.03

184.52 96

93

91.5

0 ANN

Poly (n=2)

Poly (n=3)

Poly (n=4)

Poly (n=5)

RBF (p1=5)

Fig. 13 The pictorial representation for classification performances of ANN and SVM Kernel functions. classification rate but also by the less execution time. Co-occurrence features are producing 96.0% classification rate in our scene classification problems with radial basis kernel function. This work can be further extended to classify other natural scene categories using various feature extraction methodologies. This complete work is implemented using Matlab 6.5 and SVM Toolbox.

8. Discussion This paper discusses texture based scene classification problem using Neural Networks and Support Vector Machines. Backpropagation algorithm is used for the paradigm of neural networks. Polynomial and radial basis functions are used as the Kernel functions for the support vector machine classifier. The sample images are taken from Ponce Research Group [24]. Eight Haralick features are extracted from each of the four directions for the classification. Features are extracted from all the scene categories, from raw images, without any preprocessing steps to make the system robust to real scene environments. The results show that Radial basis kernel function with SVM is giving the maximum of 96.0% classification rate in 197.03 seconds, but ANN is giving 92.5% classification rate in 770.47 seconds. The time consumption is minimal in support vector machines than artificial neural networks. In particular, radial basis kernel function is performing better than the polynomial kernel functions in support vector machines. The comparative results of ANN and SVM in detail is given in Table 2.

10. Acknowledgements The authors are grateful to the Management and Principal of Bannari Amman Institute of Technology, Sathyamangalam for their constant support and encouragement. Table 2. Comparative Results of ANN and SVM

9. Conclusion This paper concentrates on the categorization of images as ‘MIT-street’ scenes or ‘MIT-highways’ scenes. The results are proving that support vector machine is performing better than neural networks in the scene categorization problems in terms of not only by the 51

Classification Execution Time Rate (in Sec)

Classifier

TP TN FP FN

ANN

93 07 92 08

92.5 %

770.47

SVM (RBF)

97 03 95 05

96.0 %

197.03

ICGST-GVIP, ISSN 1687-398X, Volume (8), Issue (IV), December 2008

[16] G. Y. Chen and P. Bhattacharya, “Invariant Texture Classification using Ridgelet Packets”, Proc. of 18th Intl. Conf. on Pattern Recognition, 2006. [17] Devendran V et al., “ANN and SVM based Image classification using Wavelet Decomposition”, Asian Journal of Information Technology 6 (11): 1174 – 1180, 2007. [18] Devendran V et al., “Scene Categorization using Invariant Moments and Neural Networks”, IEEE Computer Society Press, Vol. 1, pp. 164-168, 2007. [19] Devendran V et al., “Invariant Moments to Scene Categorization using Support Vector Machines”, International Journal of Soft Computing 3(2): 128-133, 2008. [20] Devendran V et al., “Feature Selection for Scene Categorization using Support Vector Machines”, International Congress on Image and Signal Processing , China, 27-31 May 2008. (Accepted) [21] B.Yegnanarayana, “Artificial Neural Networks”, Prentice-Hall of India, New Delhi, 1999. [22] Vapnik, V.N., “The support vector method of function estimation. In: Generalization in Neural Network and Machine Learning. Springer-Verlag, New York, NY, pp. 239-268, 1998. [23] Burges, C.J.C., “A tutorial on support vector machines for pattern recognition”, data mining and knowledge discovery 2(2), 121-167, 1998. Available from http://svm.first.gmd.de [24] www-cvr.ai.uiuc.edu/ponce_grp/data

11. References [1] J. Henderson, “Introduction to real-world scene perception”, 12: 849-851 (3), August 2005. [2] A. Delorme, G. Richard, M. Fabre-Thorpe, “Rapid Processing of complex natural scenes: A role for the magnocellular visual pathways?”, Neuro computing 26-27, 663-570, 1999. [3] Tom Drummed, “Learning Task-specific Object Recognition and scene understanding”, Computer vision and Image Understanding 80, 315-348, 2000. [4] A. Guerin-Dugue and A. Oliva, “Classification of scene photographs from local orientations features”, Pattern Recognition Letters 21 (2000) 1135-1140. [5] A. Chella, M. Frixione and S. Gaglio, “Understanding dynamic scenes”, Artificial Intelligence 123 (2000) 89-132. [6] Matti Pietikainen, Tomi Nurmela, Topi Maenpaa, Markus Turtinen, ”View-based recognition of realworld textures”, Pattern Recognition 37 (2004) 313-323. [7] Laura Walker Renninger and Jitendra Malik, “What is scene identification just texture recognition”, vision research 44, 2301-2311, 2004. [8] Matthew Boutell and Jiebo Luo, “Beyond pixels: Exploiting camera metadata for photo classification”, Pattern Recognition 38 (2005) 935946. [9] Ian Stefan Martin, “Robust Learning and Segmentation for scene Understanding”, PhD Thesis, Dept. of Electrical Engineering and Computer science, Massachusetts Institute of Technology, May 2005. [10] Manuele Bicego, Marco Cristani and Vittorio Murino, “Unsupervised scene analysis: A hidden markov model approach”, Computer vision and image understanding 102 (2006) 22-41. [11] A. Bosch, X. Munoz and J. Freixenet, “Segmentation and description of natural outdoor scenes”, Image and Vision computing 25 (2007) 727-740. [12] Andrew Payne and Sameer Singh, “Indoor vs. outdoor scene classification in digital photographs”, Pattern Recognition 38 (2005) 1533-1545. [13] Lance M. Kaplan, “Extended Fractal Analysis for Texture Classification and Segmentation”, IEEE Transactions on Image Processing, Vol. 8, No. 11, November 1999. [14] Arivazhagan S et al., “Texture Classification using Ridgelet Transform”, Proc. of Sixth Intl. Conf. on Computational Intelligence and Multimedia Applications, 2005. [15] Arivazhagan S et al., “Texture Classification using Curvelet Statistical and Co-occurrence Features”, Proc. of 18th Intl. Conf. on Pattern Recognition, 2006.

Bibliography Devendran V started his teaching profession from April 2001 as a Lecturer in the Department of Computer Applications, Bannari Amman Institute of Technology, Sathyamangalam, Erode District, Tamilnadu, India. Currently he is working as a Senior Lecturer in the same institute from December 2006. In 1997, he has finished his Bachelor of Science in Mathematics in Vivekananda College, Madurai Kamaraj University, Madurai. He has finished his Master of Computer Applications in Ayya Nadar Janaki Ammal College, Sivakasi in the year 2000. Currently he is pursuing his doctoral programme in National Institute of Technology, Trichirappalli, Tamilnadu, India. He has worked as a co-investigator in a research project funded by DRDO, Newdelhi, India during 2003 and 2005. He has published three research papers in refereed journals. His research interest includes Object Recognition, Pattern Recognition, Texture Analysis, Neural Networks and Support Vector Machines.

52