VIEWS: 145 PAGES: 7 CATEGORY: Emerging Technologies POSTED ON: 9/5/2010
IJCSIS invites authors to submit their original and unpublished work that communicates current research on information assurance and security regarding both the theoretical and methodological aspects, as well as various applications in solving real world information security problems.
(IJCSIS) International Journal of Computer Science and Information Security, Vol. 8, No. 5, 2010 Hybrid Model of Texture Classification using 2D Discrete Wavelet Transform and Probablistic Neural Network Reem Abd El-Salam El-Deeb Taher. Hamza Elsayed Radwan Department of Computer Science, Department of Computer Science, Department of Computer Science, Faculty of Computer and Faculty of Computer and Faculty of Computer and Information Sciences, Information Sciences, Information Sciences, Mansoura University, Egypt, Mansoura University, Egypt, Mansoura University, Egypt, P.O.BOX:35516 P.O.BOX:35516 P.O.BOX:35516 Taher_hamza@yahoo.com elsfradwan@yahoo.com Reemm_db@yahoo.com Abstract— In this paper, we present a combinational approach texture classification problem based on various types of for texture classification. The proposed method analyzes texture features and different methods of feature extraction. Most of by 2D Discrete Wavelet Transforms (DWT); wavelet energy and the textural features are generally obtained from the application some statistical features construct the features vector that of a local operator, statistical analysis, or measurement in a characterizes texture. For improving accuracy the Probabilistic transformed domain [14]. Generally, the features are estimated Neural Network (PNN), which is considered as a good estimator to the probability density function, is used as a classifier that from Law’s texture energy measures, Markov random field maps input features vectors to the most appropriate texture models, Gibbs distribution models and local linear transforms classes. Two comparative evaluations have been done in order to were found not to be robust enough to allow one-to-one ensure the effectiveness and efficiency of this model. mapping between patterns and parameter sets for many reasons: the parameters computed rely on the model assumed, Keywords- Texture classification, feature extraction, discrete the neighborhoods used must not be self-contradictory and they wavelet transform, probabilistic neural network rely on the number of samples available for each combination I. INTRODUCTION of neighborhoods. In short, no model fits the observed textures perfectly, and so no model parameters are perfect in capturing Texture is the variation of data at scales smaller than the all characteristics of a texture image [7][8]. Other studies used scales of interest [7]. Techniques for the analysis of texture in Fourier transform domain, fractals and co-occurrence matrices. digital images are essential to a range of applications in areas The co-occurrence features such as contrast, homogeneity etc., as diverse as robotics, medicine and the geo-sciences. In were found to be the best of these features that they are popular biological vision, texture is an important cue allowing humans due to the perceptual meaning they have. However, they are to discriminate objects. This is because the brain is able to not adequate for texture and object discrimination as they decipher important variations in data at scales smaller than throw away most of the information conveyed by the co- those of the viewed objects. Texture may be important as well occurrence matrices [7]. in object recognition as it tells us something about the material from which the object is made. In order to deal with texture in In the recent years, wavelet analysis has become a powerful digital data, many techniques have been developed by image tool for multi-resolution analysis. Discrete Wavelet transform processing researchers [14] [7]. (DWT) and Gabor Transform are extensively used for texture Texture classification aims to assign texture labels to unknown analysis. While the DWT uses fixed filter parameters for textures, according to training samples and classification rules image decomposition across scales, the Gabor Transform by finding the best matched category for the given texture requires proper tuning of filter parameters for different scales among existing textures. Two major issues are critical for of decomposition. Further, Wavelet based methods are shown texture classification: the texture feature extraction and texture to be efficient in detection, classification and segmentation for classification algorithms [4]. many reasons: the wavelet transform is able to de-correlate the Texture feature extraction is considered as the main base of the data and achieve the same goal as the linear transformation, it efficiency of the texture classification algorithm. In order to provides orientation sensitive information which is essential in design an effective algorithm for texture classification, it is texture analysis and the computational complexity is essential to find a set of texture features with good significantly reduced by considering the wavelet discriminating power. Unfortunately, because of scale decomposition [11][14]. dependency of texture, its feature extraction has become a As denoted before the efficiency of any classification difficult problem. There have been many studies in solving system depends on effective characterization as well as 148 http://sites.google.com/site/ijcsis/ ISSN 1947-5500 (IJCSIS) International Journal of Computer Science and Information Security, Vol. 8, No. 5, 2010 choosing the appropriate classifier. Some classifier continues until some final scale is reached. The values or algorithms such as support vector machines are used in transformed coefficients in approximation and detail images some works which faced some problems because of high (sub-band images) are the essential features, which are shown algorithmic complexity and extensive memory requirements here as useful for texture analysis and discrimination. As [16]; the distance classifier is also used for measurement of textures have non-uniform pixel value variations, they can be similarity and consequent labeling but it suffered from some characterized by the values in the sub-band images or their limitation in speed and adding parameters may cause the combinations or derived features from these bands [11]. classifier to fail. [4]. In this study, a hybrid model based on the combinational LL2 HL2 approach is proposed, which combine 2D Discrete Wavelet LL1 HL1 2 HL1 Transform (DWT) and Probabilistic Neural Network LH2 HH2 (PNN) for solving texture classification problem. In the hybrid configuration, the 2D DWT is used for texture analysis and constructing features vector that characterizes LH1 HH1 LH1 HH1 the texture image by capturing all essential information. The obtained features vectors are then fed into the PNN which is used as a good estimator to probability density (a) One-Level (b) Two-Level function that help in mapping each texture feature vector to the best appropriate class with fast and efficient Figure 1. Image decomposition performance. For illustrating the effectiveness of this B. Energy model, two comparative evaluations have been done. The first one was among variety of wavelet filters for finding Energy is one of the most commonly used features for the best features extractor that provides the best texture analysis [4]. Wavelet energy reflects the distribution characterization. The other was between the PNN and of energy along the frequency axis over scale and orientation Backpropagation Neural Network (NN) as a classifier and has proven to be very powerful for texture classification. according to the mean success rates. The energy of sub-band ���� containing N coefficients is This paper is organized as follows; in section II, Discrete defined as in equation (1) [5], 1 2 Wavelet transform (DWT), Probabilistic Neural Network ���� ������������������������ = ∙ ����,���� ���� �������� , �������� 9 (1) (PNN) and Wavelet energy are mentioned. The hybrid ���� model of 2D DWT and PNN is described in section III. C. Probabilistic neural network The effectiveness of the proposed hybrid model for It is shown that, by replacing the Sigmoid activation function classification of texture images and comparative often used in neural networks with an exponential function, a evaluations are demonstrated in section IV .Finally, neural network can be formed which computes nonlinear section V presents discussion and conclusion. decision boundaries. The resulted network is considered as II. PRELIMINARIES an estimator to the probability density functions which can be used to map input patterns to output patterns and to A. Discrete wavelet transform classify patterns. This technique yields decision surfaces Wavelets are functions that satisfy certain mathematical which approach the Bayes optimal under certain conditions requirements. They are used to cut up data into different [3]. frequency components and then study each component with a PNN is a kind of these networks that called radial basis resolution matched to its scale. The basic idea of the wavelet network. It is an artificial neural network with radial basis transform is to represent any arbitrary function as a function (RBF) as a transfer function. RBF is a bell shape superposition of wavelets. Any such superposition function that scales variable nonlinearly [15]. This network decomposes the given function into different scale levels provides a general solution to pattern classification problems where each level is further decomposed with a resolution by following an approach developed in statistics, called adapted to that level [12]. Bayesian classifiers [6]. PNN is suitable for these kinds of By applying DWT, the image is actually divided i.e., classification problems for many advantages: Its training decomposed into four sub-bands and critically sub sampled as speed is many times faster than standard feed forward shown in Figure 1. (a). These four sub-bands arise from backprobagation network, it can approach a Bayes optimal separable applications of vertical and horizontal filters. The result under certain easily met conditions and it is robust to sub-bands labeled LH1, HL1 and HH1 represent the finest noise examples. scale wavelet coefficients, i.e., detail images while the sub The most important advantage of PNN is that training is easy band LL1 corresponds to coarse level coefficients, i.e., and instantaneous that weights are not “trained” but approximation image. To obtain the next coarse level of assigned. Existing weights will never be alternated but only wavelet coefficients, the sub band LL1 alone is further new vectors are inserted into weight matrices when training. decomposed and critically sampled. This result in two- level So, it can be used in real-time. Since the training and running wavelet decomposition as shown in Figure 1. (b).This process 149 http://sites.google.com/site/ijcsis/ ISSN 1947-5500 9 9ال (IJCSIS) International Journal of Computer Science and Information Security, Vol. 8, No. 5, 2010 procedure can be implemented by matrix manipulation, the speed of PNN is very fast [15]. The probabilistic neural network uses a supervised training set to develop distribution functions within a pattern layer. These functions, in the recall mode, are used to estimate the likelihood of an input feature vector being part of a learned category, or class. The learned patterns can also be combined, or weighted, with the a priori probability, also called the relative frequency, of each category to determine the most likely class for a given input vector. If the relative frequency of the categories is unknown, then all categories can be assumed to be equally likely and the determination of category is solely based on the closeness of the input feature Figure 2. PNN Architecture vector to the distribution function of a class [6][1]. The bias b allows the sensitivity of the radial basis neuron to Probabilistic neural networks can be used for classification be adjusted. Each bias in the first layer is set to problems. When an input is presented, the first layer 0.8326/SPREAD. This determines the width of an area in the computes distances from the input vector to the training input space to which each neuron responds. SPREAD should input vectors and produces a vector whose elements indicate be large enough that neurons respond strongly to how close the input is to a training input. The second layer overlapping regions of the input space [13]. sums these contributions for each class of inputs to produce as its net output a vector of probabilities. Finally, a competed The Probabilistic Neural Network is based on Bayesian transfer function on the output of the second layer picks the classification and the estimation of probability density maximum of these probabilities, and produces a 1 for that function that is necessary to classify the input vectors into class and a 0 for the other classes. The architecture for this one of the target classes approaching the Bayesian optimality system is shown below in Figure 2. [6]. [15]. It is assumed that there are ���� input vector/target vector pairs. Each target vector has ���� elements. One of these elements is III. HYBRID MODEL FOR TEXTURE CLASSIFICATION 1 and the rest are 0. Thus, each input vector is associated The texture classification scheme is based on two with one of ���� classes. principles, choosing features that provide the best The first-layer input weights ��������1 are set to the transpose of characterization to the texture image and working with fast, the matrix formed from the ���� training pairs, ����′. When an easy and robust classifier in order to reach the best input is presented, the || ���������������� || box produces a vector whose classification result. elements indicate how close the input is to the vectors of the In this study a viable algorithm with high precision and training set. These elements are multiplied, element by low calculating load is proposed to classify texture images element, by the bias and sent to the radial basis transfer using wavelet transform and its combination with function. An input vector close to a training vector is probabilistic neural network. In the proposed combinatory represented by a number close to 1 in the output vector a1. If configuration the DWT and PNN function as black boxes an input is close to several training vectors of a single class, in a complementary manner. The functionality manner it is represented by several elements of a1 that are close to 1. involved can be combined in two phases: The second-layer weights ��������2 are set to the matrix T of target vectors. Each vector has a 1 only in the row associated i. Texture characterization phase and with that particular class of input, and 0’s elsewhere. The ii. PNN classification phase. multiplication Ta1 sums the elements of a1 due to each of the ���� input classes. Finally, the second-layer transfer function, The texture classification phase starts with taking the texture compete, produces a 1 corresponding to the largest element images as an input and with the help of the DWT, the texture of n2, and 0’s elsewhere. Thus, the network classifies the images are analyzed and features vectors are constructed. input vector into a specific ���� class because that class has the The obtained features vectors are entered to the PNN for maximum probability of being correct [9]. training which starting the PNN classification phase that continues with testing and ends with displaying the classification result as illustrated in Figure 3. 150 http://sites.google.com/site/ijcsis/ ISSN 1947-5500 (IJCSIS) International Journal of Computer Science and Information Security, Vol. 8, No. 5, 2010 calculate the standard average of matrix elements that is computed as in equation (2), 1 Texture image ���� = ∙ ���� |����(�������� , �������� )| ����,���� (2) ���� And the last four elements represent the standard deviation of DWT the approximation and the detail coefficients matrices which measure the variability in matrix elements and are computed as the square root of variance as in equation (3), Calculate wavelet energy and wavelet 1 ���� 2 statistical features for 1-level wavelet ���� = ����=1 �������� − ���� (3) ����−1 decomposed sub-bands. Wavelet Energy Mean Standard Deviation LL LH HL HH LL LH HL HH LL LH HL HH Features vectors Figure 4. Features vector Hence, at the end of this phase, each texture image has been Train PNN well characterized by 12 elements features vector that captures all essential information needed for discrimination. B. PNN classification phase Test PNN PNN is considered as a good estimator to probability density function that can be used to map input patterns to output patterns and to classify patterns efficiently with fast Display and compare the result execution and ease implementation. PNN classification phase has two parts, training and testing. After constructing the features vectors that represent the Figure 3. Structure of Hybrid Classification model texture images in the texture characterization phase, the network is trained with the features vectors and the corresponding texture images as described in the following A. Texture characterization phase steps: In order to overcome the obstacle of texture characterization due to its scale dependent property, the 1) The input is of size ���� × ���� with feature elements discrete wavelet transform is used as a powerful tool for multi- ���� = 12 and training samples����. resolution analysis. 2) Radial basis layer weight ��������1 is set to the transpose For wavelet decomposition of various texture images, the of the ���� × ���� matrix of training samples that ��������1 is of decomposition at 1-level is performed using different size���� × 12. wavelet transform filters. Thus, the image is decomposed 3) The dot product between the input vector ���� with size into one approximate image with one approximation 12 × 1and the ���� ����ℎ row of ��������1 produces the ���� ����ℎ coefficients LL and three detail images with horizontal LH, element of the distance vector ||��������1 − ����|| whose vertical HL and diagonal HH detail coefficients. size���� × 1. Wavelet transform of an image measures light fluctuation in 4) The radial basis layer biases ���� are all set to0.8326 ∕ different scales. Therefore, the wavelet energy that reflects ������������������������, that ������������������������ is a constant chosen the distribution of energy along the frequency axis over scale according to experiment. and orientation is calculated for the approximation and the 5) The net input ����1 is obtained from element-by-element detail coefficients matrices. Also to increase sensitivity and multiplication of the bias vector ���� with the distance precision, some wavelet statistical features are calculated vector ||��������1 − ����||that denoted as such as the mean and the standard deviation of the approximation matrix as well as the mean and the standard ����1 = ||��������1 − ����|| ∙∗ ���� (4) deviation of the detail coefficients matrices; and then they are added to image features. The features obtained construct 6) The transfer function is the radial basis function that a feature vector with 12 elements organized as illustrated in Figure 4. The first four elements represent wavelet energy of defined as in equation (5) and its shape is illustrated the approximation and the detail coefficients matrices which in Figure 5. is computed as in equation (1). 2 The second four elements represent the arithmetic mean of the ���� = ������������������������ ���� = ���� −���� (5) approximation and the detail coefficients matrices which 151 http://sites.google.com/site/ijcsis/ ISSN 1947-5500 (IJCSIS) International Journal of Computer Science and Information Security, Vol. 8, No. 5, 2010 success rates for each classifier. The obtained results are discussed below. The number of incorrect classifications of the PNN to the tested texture images using the wavelet filters mentioned before is tabulated in TABLE I. It is found that the wavelet filters db2, bior2.2 and db3 give the best characterization to texture images Bark.0000 and Paintings.11.0003, wavelet filters sym6, bior2.2 and coif2 give the best characterization to leaves.0008, wavelet filters db2, sym6 and bior2.2 give the best characterization to Metal.0004 and wavelet filters sym6, Figure 5. Radial basis function db5, coif2 and db3 give the best characterization to Flowers.0006. If the input ���� is identical to the ���� ����ℎ row of ��������1 , then the TABLE I.THE NUMBER OF INCORRECT CLASSIFICATIONS WITH VARIOUS ���� ����ℎ element of ����1 is equal to 1 WAVELET FEATURES If the input ���� is close to the ���� ����ℎ row of ��������1 , then the radial basis function produces a value near 1, else it produces a value Number of incorrect classifications far from 1. No Texture images 7) Competitive layer weights matrix ��������2 is set to (db2) (sym6) (db5) (bior2.2) (coif2) (db3) ���� × ���� matrix with target classes����. 1 Bark.0000.pgm 1 3 2 1 3 1 8) In competitive layer, the vector ����1 is multiplied with 2 Bark.0006.pgm 1 1 1 1 1 1 matrix ��������2 producing the output vector ����2 of size 3 Fabric.0015.pgm 1 1 1 1 1 1 ���� × 1. 4 Flowers.0006.pgm 2 1 1 2 1 1 9) The competitive function ���� produces 1 corresponding 5 Food.0000.pgm 0 0 0 0 0 0 to the largest value of ����2 and 0 elsewhere. 6 Leaves.0008.pgm 1 0 1 0 0 2 7 Metal.0002.pgm 0 0 0 0 0 0 10) For testing the network, an unknown features vector is 8 Metal.0004.pgm 0 0 1 0 1 2 entered as input and the network classifies it according 9 Misc.0003.pgm 1 1 1 1 1 1 to the class associated with the largest probability. 10 Paintings.11.0003.pgm 2 4 3 2 4 2 11 Water.0001.pgm 0 0 0 0 0 0 12 Stone.0004.pgm 1 1 1 1 1 1 IV. EXPERIMENTAL RESULT 13 Wood.0002.pgm 0 0 0 0 0 0 14 Water.0005.pgm 0 0 0 0 0 0 In order to assess the discrimination capability of the proposed hybrid model, experiments are conducted with 14 Vision Texture (Vistex) images database, each of size 512×512 which The mean success rates of the chosen wavelet filters are shown is divided into sub-images of size 64×64 for totally 896 texture in Figure 6. It is found that the characterization of texture images with 14 texture classes. images using bior2.2 wavelet filter has the highest mean First, for applying texture characterization phase, one-level success rate (93.57%) and the lowest mean success rate wavelet decomposition is applied to each texture image using (90.72%) was to the characterization using coif2 wavelet filter db2, sym6, db5, bior2.2, coif2, db3 wavelet filters separately. which is observed in Figure 7. After extracting the features, each texture image is characterized and represented with 12 elements features vector The Mean success rates (%) constructing totally 896×12 features vectors for all texture images used. Then, before entering the second phase, the PNN 93.57% classification phase, the features vectors obtained at the end of 92.86% the texture characterization phase are divided into 756 features vectors for training and 140 for testing. Finally, the PNN is trained with 756×12 input features vectors 91.43% 91.43% 91.43% and 14 target classes. After training, the PNN is tested with 90.72% 140×12 features vectors. For illustrating the effectiveness and efficiency of this model, two comparative evaluations have been done. The first one was between features extracted by the chosen wavelet filters according to the corresponding success rate in order to evaluate the efficiency of characterization for each db3 coif2 bior2.2 db5 sym6 db2 filter. And the other was between the PNN and Backpropagation NN as a classifier according to the mean Figure 6. The mean success rates of various wavelet filters 152 http://sites.google.com/site/ijcsis/ ISSN 1947-5500 (IJCSIS) International Journal of Computer Science and Information Security, Vol. 8, No. 5, 2010 100 V. CONCLUSION 90 80 Texture classification problem has become the business of a 70 lot due to its great effect on many fields. As an attempt to 60 solve this problem many studies have been done in spite of 50 40 hindrances they suffered in the two sides, characterization and 30 classification. As for characterization, researchers tried to 20 10 choose features that provide the best representation of the 0 texture images in the presence of texture scale dependent property. While they work on the use of the appropriate classifier algorithm to provide the best discrimination capability. In this work a hybrid model is presented to classify texture bior2.2 coif2 images. The 2D DWT is combined with the PNN constructing this hybrid model. The discrete wavelet transform is used as a Figure 7. Comparison between the best and the worst characterization powerful tool for multi-resolution analysis so it is used for according to the corresponding correct classification rates of 14 texture texture analysis as an attempt to overcome the obstacle of classes. texture scale dependent property. The PNN is a radial basis network that is considered as an estimator to the probability The features vectors that have been obtained from bior2.2 density functions which can be used to map input patterns to wavelet filter which represent the best characterization of the output patterns and to classify patterns. The PNN is suitable texture images have been used to compare the performance of for these kinds of classification problems that it can approach the PNN used in this model and the Backpropagation NN in a Bayes optimal result under certain easily met conditions; as classification. well as the training is easy, fast and robust. As illustrated in Figure 8, the mean success rate that has been The structure of the proposed hybrid model is divided into two obtained over the 14 texture classes using the PNN is far more phases. The first phase is the texture classification phase in than the other obtained using the Backpropagation NN. In which 1-level wavelet decomposition has been performed and addition to the great differences in speed and simplicity of the 4 sub-bands have been obtained representing the performance between the two classifiers prove the approximation, vertical, horizontal and diagonal detail. As a effectiveness and efficiency of the PNN that has been used in way to select features that capture all the essential information this model with respect to Backprobagation NN. needed to uniquely characterize the texture, wavelet energy that reflects the distribution of energy along the frequency axis The Mean Success Rates of PNN vs. NN (%) over scale and orientation has been calculated for the approximation and the detail coefficients matrices. Also to 100 increase sensitivity and precision, some wavelet statistical 90 features are calculated such as the mean and the standard 80 deviation of the approximation matrix as well as the mean and 70 the standard deviation of the detail coefficients matrices; and 60 then they are added to image features. The features obtained 50 construct a feature vector with 12 elements which is fed with 40 the corresponding target class as input to the PNN starting the 30 second phase, the PNN classification phase. In the PNN 20 classification phase, the PNN is trained with the input features 10 0 vectors then it is tested with other features vectors for evaluating its discrimination capability. Experiments have been conducted for evaluating the performance of the proposed hybrid model. The model proved that the features derived from the approximation and detail coefficients, the wavelet energy and the statistical features, uniquely characterize a texture. In order to find the best Probablistic Neural Network Neural Network features extractor a comparative evaluation has been done with features extracted by different wavelet filters and the Figure 8. Comparison between the mean success rates that have been corresponding correct classification rates. Another obtained using the PNN verses the mean success rates that have been comparative evaluation with respect to classifiers has been obtained using Backpropagation NN. performed between the PNN and Backpropagation NN which provided great evidence about the effectiveness of the PNN as 153 http://sites.google.com/site/ijcsis/ ISSN 1947-5500 (IJCSIS) International Journal of Computer Science and Information Security, Vol. 8, No. 5, 2010 a classifier for texture images with simple structure, fast [5] G.Van de Wouwer, P.Scheunders and D.Van Dyck, “Statistical texture characterization from discrete wavelet representations,” IEEE execution and high performance. Transactions on Image Processing, Belgium,1999 This hybrid classification model achieved a good progress in [6] Leila Fallah Araghi, Hamid Khaloozade and Mohammad Reza Arvan, solving the texture classification problem by the proposed “Ship Identification Using Probabilistic Neural Networks (PNN),” The complementary manner of the 2D DWT and PPN. Compared International Multi-Conference of Engineers and Computer Scientists, with other methods, the system functionality is fast in Hong Kong, March 2009 execution, efficient in recognition and easy in implementation. [7] Maria Petrou and Pedro Garcia Sevilla, “Image processing-Dealing with texture,” John Wiley&Sons Ltd, England, 2006 We are still in a need of finding more features that capture [8] M.Ghazvini, S.A.Monadjemi, N.Movahhedinia and K.Jamshidi, more essential information of texture in order to provide the “Defect Detection of Tiles Using 2D-Wavelet Transform and Statistical best characterization and achieve the optimal classification Features, ” World Academy of Science,Engineering and Technology, results. More extended efforts are under development in order Iran, 2009 to improve the efficiency of the system. [9] Neural Network Matlab Toolbox 2008 [10] Sami Gazzah and Najoua Essoukri Ben Amara, “Arabic Handwriting ACKNOWLEDGMENT Texture Analysis for Writer Identification Using the DWT-Lifting Scheme, ” Ninth International Conference on Document Analysis and Recognition, Parana, 2007 First of all, I thank my father, my mother, my sisters and [11] S.Arivazhagan and L. Ganesan “Texture Classification using Wavelet Transform ,” The Sixth International Conference on Computational my brother for their moral support I required in my life at all. Intelligence and Multimedia Applications (ICCIMA’05), IEEE, India, I deeply thank my advisor, Dr. ElSayed Radwan, whose help, 2005 advice and supervision was invaluable. [12] S.Arivazhagan, S.Deivalakshmi and K.Kannan, “ Performance Analysis of Image Denoising System for different levels of Wavelet Lastly, I offer my regards and blessings to my friends and all of decomposition,” International Journal of Imaging Science and Engineering (IJISE), USA, JULY 2007 those who supported me in any respect during the completion of the paper. [13] S.N. Sivandm, S. Sumathi and S.N.Deepa, “Introduction to Neural Netwrks Using Matlab7.0,” Tata mcgraw-Hill Publishing Company Limited, Copyright 2006. REFERENCES [14] S.Shivashankar and P.S.Hiremath, “Wavelet Based Features For Texture [1] Ali H. Al-Timemy, Fawzi M. Al-Naima and Nebras H. Qaeeb, Classification,” ICGST International Journal on Graphics, Vision and “Probabilistic Neural Network for Breast Biopsy Classification,” Image Processing, India, December 2006 MASAUM Journal of Computing, Iraq, September 2009. [15] Stephen Gang Wu, Forrest Sheng Bao, Eric You Xu, Yu-Xuan Wang, [2] Anagnostopoulos I., Anagnostopoulos C., Vergados D., Loumos V., and Yi-Fan Chang and Qiao-Liang Xiang, “A Leaf Recognition Algorithm Kayafas E., “A Probabilistic Neural Network for Human Face for Plant Classification Using Probabilistic Neural Network,” arXiv, Identification based on Fuzzy Logic Chromatic Rules,” IEEE MED03, USA, July 2007. Greece, 2007. [16] http://www.svms.org/disadvantages.html, 2010-07-07 [3] Donald F. Specht, “Probabilistic neural networks for classification, mapping, or associative memory,” IEEE International Conference on Neural Networks, California , 2001. [4] Engin Avci, Abdulkadir Sengur and Davut Hanbay, “An optimum feature extraction method for texture classification,” Pergamon Press, Inc. Tarrytown, NY, USA, 2009. 154 http://sites.google.com/site/ijcsis/ ISSN 1947-5500