Texture Images Classification Using Support Vector Machine

The purpose of the research of content-based image retrieval (CBIR) is to realize retrieving images automatically and intelligently. The objects of the research are methods and technology that can help the user retrieve particular images from image database conveniently, quickly and accurately. Texture feature classification is the key technology in the texture image CBIR system. The traditional statistical predictive methods require transcendent knowledge. The researching Premise of them is that the number of samples is infinite.When the dimension increases, the performance of texture classification sometimes decline. The use of these methods in texture classification is limited because when doing the nonlinear texture classification, the texture feature should be mapped to a higher dimension. So a method influenced by the dimension little is needed to carry out the nonlinear texture classification.Before classifying texture feature, we must first extract texture feature. This thesis adopted the method of Gabor Wavelet to extract texture feature. In the past, when constructing Gabor filters, there was not a uniform method to select parameters. Different authors used different parameters because the effects of parameters are not clear. This paper researched the effects of parameters of Gabor function in texture classification and found out the importance of them by carefully designed experiment. The rusult can be used to guide the construction of Gabor filters. Based on Linear Predictive Coding and Gauss function, this paper constructed a new kernel function, and analysed the separable and local character of the kernel function. We got the conclusion that the kernel function has good linear separable character and interpolation ability, and is good at extacting the local feature. Finally, we took the advantage of Polynomial kernel function in extracting overall feature, and combine the new kernel function and the Polynomial kernel function to get a final kernel function and a SVM, LG-SVM.In the experiment of texture classification, we compared the new method with two classical methods. The results showed that the classification precision of our method was always higher than others' and as the dimension increases, the performance of other method dropped sometimes, but our method was influnced little.