Study Of Texture Feature Extraction And Classification Method

As an important part of pattern recognition and computer vision, textureclassification has been widely used in many fields. The extraction methods of texturefeatures and the classifier are most important to the classification result. For uniformgray texture images, with wavelet transform, local binary pattern and other texturefeatures extraction methods, we can extract the texture features, and then classify thetexture images with the Neural Network and Support Vector Machine (SVM).However, for the color texture images and the non-uniform texture images, theclassification accuracy still need to be improved since the existing methods effects arenot accurate enough.In view of the current existing problems of texture image classification, thispaper presented a new approach to extract texture features of color texture images andnon-uniform light distribution texture images, and then performed some experimentsimulation by combining with SVM. The main contents and contributions of thedissert are as follows:(1) This paper presented a new approach for automatically classify the vignettingtexture images. By extracting the texture features with the wavelet packetdecomposition algorithm, the vignetting coefficients were utilized to adjust the gottenwavelet packet coefficients, then the effect of vignetting to the texture features waseliminated, and consequently the texture classification accuracy was greatly improved.Experimental results show that the approach in this paper can improve theclassification accuracy significantly, and achieve ideal texture classification effect.(2) This paper presented two approaches to extract the color texture’s features.One is to extract the color co-occurrence matrix between each layer of color imagesbased on the idea of statistics. Another one is to build the texture index based on theoriginal texture primitives. Statistic the original texture primitives as local texturefeature, and extract hue and saturation components of the color image as globaltexture feature. Experimental results show that by combining with SVM, theapproaches in this paper can improve the classification accuracy significantly.In addition, this paper discussed the statistical land feature method, andperformed some simulation experiment with the gray texture images in the imagedatabase. Experimental results show that although this method can classify texture images automatically, its time consumption is high. This paper also introduced thetexture feature to the image inpainting field, the size of patches was automaticallyselected according to the relationship with the extracted texture features. Simulationresults show that the perfect effect of texture synthesis can be obtained with thepatches size selected by the proposed method.