Texture Classification Based On Coefficient Co-occurrence Matrix And Improved Multi-class SVM

Texture analysis is very important in pattern recognition. There are many methods in texture research, such as statistical method, modeling, and spectral method. In this paper, spectral method and co-occurrence matrix in statistical method are described in detail. Wavelet which has a characteristic of multi-resolution is in accordance with the visual multi-channel of human, and is widely used to describe texture. After the discrete wavelet decomposition of each layer there is an step of sub-sampling, if input image change slightly, then there will be a large change of wavelet coefficient, especially in large-scale of decomposition, the change is more evident. Discrete wavelet flame transform with the characteristic of shift invariance removes the process of sampling, and is more suitable for texture image processing.In addition, the multi-classification algorithm based on support vector machine is studied in this paper. Recently support vector machine based on binary tree is widely used. The core of the algorithm is the structure of the binary tree, the common methods are clustering distance, spatial distribution and clustering method. In the clustering method, the categories which are easy to classify will be separated firstly. This algorithm can construct various types of appropriate binary tree on the basis of samples, and it's decision-making speed and generalization ability are higher and better than the first two. Support vector machine based on Huffman tree simulates the clustering method.Through the research above, in this paper the improvements of co-occurrence matrix of coefficient and multi-classification of support vector machine are made as following:(1)In this article, after the discrete wavelet flame transform of input image, unlike the previous we extract two kinds coefficient co-occurrence matrix of both different scales" and one scale's. Finally, extract statistics of co-occurrence matrix as the texture feature vector. We choose sixteen types of texture images in the Vistex texture database. After twice discrete wavelet flame transform, we get five co-occurrence matrix, then extract statistics of co-occurrence matrix. We use one to one support vector machine to classify the image. Experiments show that after normalization, the proposed method gets higher classification accuracy than the GLCM method, and the previous coefficients co-occurrence matrix methods.(2)Support vector machine based on Huffman tree simulates the clustering method, this method reduces the amount of computation for removing the process of clustering. The distance between the classes is used as measure of different classes. In this article Euclidean distance and Mahalanobis distance are applied to construct the Huffman tree. We choose five kinds dataset in the UCI database such as glass, wine, ecoli, etc. Through experiments compared with other ways, the proposed method gets higher classification accuracy than one to one method and one to rest method, and higher speed than the clustering method for removing the step of clustering.