Abstract-Thispaper presents the comparative analysis of texture image classification usingthree methods. Texture is a repeating pattern of local variation in imageintensity.
The texture provides information in the spatial arrangement ofcolors or intensities in an image, thus texture is a feature used to partitionimages into regions of interest and to classify those region. Thecontent based image retrieval technique (CBIR) is very effective ifclassification of large scale general purpose image database into textured andnon textured images is done. A technique to accurately classify the images intotextured or non textured category is based on image features. In this paper wepresent three methods comparison purpose for classification of textured image.The third method is proposed method based on neural network method exceptingthat gives better accuracy for image classification.Keywords- Textured image, Support Vector Machines, Grey Level, Imagesegmentation, Wavelet transforms.1.Introduction-Textureclassification is important in content based image retrieval (CBIR) system.
TheCBIR is technique for retrieving semantically relevant images from an imagedatabase based on automatically derived image features. Texture classificationis concerned with identifying given textured region from given set of texturedclasses. The texture classification is basically classifying pixels in an imageaccording to their texture cues .Three principles approaches used in imageprocessing to describe the texture of region are Statistical ,Spectral, Support Vector MachinesIn this paper the following three methods were discussed1. Classification ofimage using color and texture attributes.2. Texture ImageClassification Using Support Vector Machine3.
TextureClassification Based on Neural network and wavelet Transform 2.Classification of image usingcolor and texture attributes.-In this method wepropose an algorithm to improve the accuracy of this classification byemploying wavelet transform for extraction of feature for monochrome as well ascolor images. We use an algorithm toclassify a Photographic image as textured and non textured, using regionsegmentation and statistical testing. 1The algorithm uses well known LUV color space where L encodes frequencyinformation (luminance). U and v encodes color information (chrominance).
Toobtain remaining three feature the Harr wavelet transform is applied to the Lcomponent of the image.7 The k-means algorithm is used to cluster the featurevectors into several classes with every class corresponding to one region inthe segmented image. The k-means algorithm is a well-known statisticalclassification algorithm The k-means algorithm is used to cluster the featurevectors into several classes with every class corresponding to one region inthe segmented image. K-mean algorithm uses pixel wise segmentation instead ofblock wise segmentation.3 After applying K-means clustering algorithm weobtain different classes.
To classify images into the semantic classes texturedor non-textured, a mathematical description of how evenly a region scatters inan image is the goodness of match between the distribution of the region and auniform distribution. The goodness of fit is measured by the x2 statistics.Textured and non-textured imagesare classified by thresholding the average x2 statistics for all the regions inthe image. m X2= 1/m ? xi2 2.
1 i=1 Where Xi2 —- statistics in region i (i= 1 …m) X2 — average statistics Where Xi2 = 2.2 If X 2 < 0.32, the image islabeled as textured; otherwise, non-textured.
The histograms of X 2 for the two types of images areshown in Figure. It is shown that the two histograms separate significantlyaround the decision threshold 0.32.
3Fig.2.1Textured Vs Non textured image For generalpurpose images such as images in photo library, images on www (world wide web)etc automatic image classification for CBIR is difficult. In this method awavelet transform based algorithm for computation of feature vector isproposed.
We used a general-purpose image database containing 100 images of Dr.J. Z. Wang database.
These images are pre-categorized into 10 groups: Africanpeople, beach, buildings, buses, dinosaurs, elephants, flowers, horses,mountains & glaciers, and food. All images have the size of 384×256 or256x386. All images are stored in JPEG format.1. SimulationResult Sr. no Image no X12 X22 X32 X42 X52 X62 X2 Final Result 1 0 0.0448 0.0970 0.
1111 0.2060 0 0 0.0765 Textured 2 4 0.1156 1.0572 0.
7588 0.5204 0 0 0.4086 Non Textured 3 7 0.7242 0.4817 0.2200 0.3859 0 0 0.3020 Textured 4 17 0.
8285 0.3974 1.2917 1.3316 0 0 0.
6415 Non Textured 5 19 0.9563 0.3261 0.2291 0.5496 0 0 0.3435 Non Textured Table2.
1The Image and final result by using X2 Fig2.2Sample images The imagedatabase is downloaded from website http//www.DB.Standford.edu/Image.
Furtherstatistical method is used for the classification of images into textured ornon textured classes so that the search domain for the CBIR is reduced. Thealgorithm is compared with the standard method and found to classify the imageswith good accuracyFrom this methodit is concluded that an algorithm for classification of images intotextured/Non-textured images is implemented. The limitationof this method is this algorithm classify only image and to improve thesegmentation and classification accuracy, the study of using the shape featuresinto account during pixel clustering and similarity distance computation can beconsidered.
Also it doesnot provide any information about Contrast,Correlation, Energy, and Homogeneity for texture classification .This canimprove by support vector machine (SVM) technique that is discussed in secondmethod. 3. Texture Image Classification Using SupportVector Machine- Texture is defined as a pattern that is repeated and isrepresented on the surface or structure of an object.
To separate textures intoa single texture type, first we need to preserve spatial information for eachtexture. For instance, the manual grey level thresholding which does notprovide the spatial information for each texture that could generate inappropriate segmentation result. Grey Level Co-occurrence Probabilities (GLCP)statistics are used to preserve the spatial characteristics of a texture.
Theselection of certain texture is possible based on the statistical features. Thebest statistical features that are used for analysis are entropy, contrast, andcorrelation . However, further analysis in shows that correlation was notsuitable for texture segmentation. GLCP statistics can also be used todiscriminate between two different textures. Boundaries can be created from theshift on statistical feature while moving from one texture to another.2 Support Vector Machine (SVM) is a typeof training method which is used to separate extracted features by creating aseparating hyper plane .
SVM have been proven to overcome the local minimumthat happens in Neural Networks (NN) training algorithms. Thus, SVM provides abetter performance in terms of accuracy for classification and regression. Inimaging, SVM is modified to do several classification tasks, such as pattern recognition,in edge detection, in texture classification and video classification. SVMserves as complement for image segmentation methods. 3.
1.1METHODOLOGYThis methodconsiders the problem of texture classification only for a gray-level casewhich is conventionally tackled in two stages of feature extraction andclassification. 3.
1.2GLCP Feature Extraction:GLCP is adiscrete function that represents joint probability, Cij, of different sets ofpixels having different grey levels, and is defined by 3.1 where Fij is theco-occurrence matrix constructed by the frequencies of two grey levels of tworelational pixels. G represents the grey level quantization. The distancebetween two relational pixels is set to become 1 for micro-texture analysis.The common angle is either 0°, 45°, 90° or 135°.
To reduce the computation timein GLCP feature extraction, we set a window size, M×N or a block of pixels asone feature value. 2 3.2.SVM ClassificationThe purpose ofSVM is to map feature vectors into a higher dimensional feature space, and thencreating a separating hyper plane with maximum margin to group the GLCPfeatures.
Support vectors (SVs) contain highlighted pixels that help to createthe margins or boundaries in an image. The higher dimensional space is definedby a kernel function. The kernel functions that we used in texturediscrimination are shown in Table 3.1.
Table.3.1: Kernel functions for used in SVM training 2, 93.3 Simulation Result: In this SVM methodthe dataset is 15 types of texture images which are retrieved from bordatzdatabase 13. Each type of texture image consists of 9 equal size samples.
Texture images are rice, oriented rattan, handmade paper, fur image, pressedcork, grass, straw etc. All images are stored in PNG format. The texture imagedatabase is downloaded from website:http://perso.telecom-paristech.fr/~xia/invariant_texture/invariant_texture_brodatz/Brodatz_re.htmlThesample image is shown Fig.
3.1 Sample image 1_1.pngFigure 3.2:Thegraph of GLCP statistical features generated from figure3.1In this method the texture related parameters likecontrast, correlation,enerry and homogenety are calculated and depending onthese parameter the image classification is done.From above table it is shownthat, for multiclass classification RBF(Radial basis function) provide maximum correct classification rate.
Therefore more multiclass classification kernel choose to be Radial basisfunction, it is shown in table 3.2. Table 3.
2: Experimentalresults for accuracy that can be achieved in Developed System 2 Sr No. Kernel parameter Classes Accuracy % training sample Time (Sec) 1 Linear 2 100 10 79.45 2 RBF with c=128,g 0.125 5 90 25 161.
19 Where c = set the parameter c for regularized supportvector classification, g = set gamma in kernel function. Fromtable 3.2, Experimental results shows accuracy for multiclass classification byselecting kernel RBF with c =128,g = 0.125.2Figure3.3 shows graph for Accuracy versus number of training samples per class.
Theaverage accuracy is achieved 80%. Figure3.3 Accuracy of texture classification of SVM system 2From this method it is concluded that an algorithm fortexture image classification using support vector classification is proposedand implemented.
This algorithm classifies texture images using GLCP and SVM asa feature extraction and classification. SVM can be considered as a modernclassification approach which features a lot of benefits, such as kernel trickand soft-margin classifiers.The drawback of this method is classification accuracyget reduced as training sample increases as well as execution time is alsoincreased. 4.Texture Classification Based on Neural network and wavelet TransformIn this methodneural network and discrete wavelet transform is used for classifying texturedimages.7The multi resolutionanalysis is applied to textured images to extract a set of intelligiblefeatures. These extracted features, in the form of DWT coefficient matrices,are used as inputs to four different multilayer perception (MLP) NeuralNetworks and classified.
This is proposed method for texture classification. Weexpect that higher classification accuracy can be obtain as we increasetraining sample.12NeuralNetwork as a Classifier-The feed forward neural network, and a description ofthe back propagation learning algorithm is given, which is very help full forclassification of texture.
The basic building block of an artificial neuralnetwork is the neuron. The connection weights between neurons are adjusted. Theneuron receives inputs opi from neuron ui while the network is exposed to input pattern p. Each input ismultiplied by a connection weight wij, where wij is the connection betweenneurons ui and uj. The connection weights correspond to the strength of theinfluence of each of the preceding neurons. After the inputs have been multipliedby the connection weights for input pattern p, their values are summed, net pj.Included in the summation is a bias value ?j to offset the basic level of theinput to the activation function, f (net pj), which gives the output opj.
12 Figure4.1 shows the structure of the basic neuron.Fig 4.1 Basic Neuron Anartificial neural network is a system of processing elements (PE)interconnected by various synaptic strengths Recently, they have become popularclassification devices for both one-dimensional and two-which use a gradientdescent learning algorithm called back propagation (BP) and a topology calledmultilayer perceptron (MLP) have been the most dominant structure forclassification purposes. Back propagation uses a squared error cost functionwhich expresses the difference between the actual and desired responses of thenetwork.Inthis method the features are extracted using Discrete Wavelet Transform (DWT)is proposed. The spatial-frequency information which a DWT contains is idealfor classifying such images as textures.
The fourseparablesub matrices at any given resolution level: low low(LL), LH, HL, and HH. 7Theblock diagram for proposed method is as shown in fig 4.2.12 Fig.4.2Proposed method for Texture Classification Based on Neural network and waveletTransform Fromthis proposed method we are expecting better classification rate reducingexecution time. 5.
ConclusionFromabove comparative analysis it is concluded that in first method, algorithmclassify only image as texture or non texture .It does not provide anyinformation about texture. So we use second method of SVM .
In this method algorithm classifies texture imagesusing GLCP and SVM as a feature extraction But drawback of this second methodis as training sample increases the classification accuracy of texture imageget redued so to over come this drawback we proposed the third method in whichwe will try to improve the classification accuracy by using neural network anddiscrete wavelet transform.
