Research On Feature Extraction For Texture Image Recognition

Texture images are ubiquitous in the real world,and the texture patterns contained in texture images play an important role in human's perception of the real world.Therefore,it is of great significance to do researches on the recognition of texture images,where the feature extraction of texture images is the key component.However,the existing methods of feature extraction on texture images have the following drawbacks:(1)the existing methods lack the comprehensive robustness to illumination conditions,image rotation,scale variation and the number of training samples;(2)the existing methods cannot achieve high recognition accuracy and high computational efficiency at the same time.To solve the problems mentioned above,this dissertation conducted further researches on the feature extraction of texture images,and the main innovative achievements are as follows.(1)A method based on local feature description and texton learning is proposed for feature extraction of texture images.This method constructs a descriptor composed of five characteristic components(i.e.local entropy,local variance,local range,local difference sign count,and local difference magnitude count)to describe the features of potential textons,and the K-means clustering algorithm is used to learn the texton dictionary.Experimental results show that the local descriptor adopted in this method is very discriminative and low-dimensional(i.e.seven dimensions),and the scale of the learned texton dictionary is smaller(i.e.only a half of that of traditional texton learning methods).These properties mentioned above significantly reduce the time consumption of traditional texton learning methods in local feature extraction,texton dictionary learning,texton encoding,and similarity measurement.Meanwhile,the proposed method can achieve higher recognition accuracy than traditional texton learning methods.(2)A method based on the fusion of global and local encoding Gabor features is proposed for feature extraction of texture images.This method first constructs a pyramid space with four levels for each texture image via downsampling and upsampling,and then Gabor filtering is implemented by the convolution of each image in the pyramid space with the designed Gabor filter bank that has multiple scales and orientations.After getting the filtered images,the mean and variance of magnitude images are used as global Gabor feature,and the joint encoding of magnitude and phase images is used as local Gabor feature.Finally,the fusion of global and local Gabor features,and texture image recognition are realized in the framework of nearest subspace classifier.Experimental results show that the proposed method can extract discriminative texture features,and significantly outperforms traditional Gabor filtering methods in terms of recognition accuracy.Meanwhile,the proposed method is efficient and robust to scale variations and the number of training samples.(3)An improved CLBP algorithm with robustness to illumination,rotation,and scale variations is proposed for feature extraction of texture images.This method uses the dominant direction of local pattern to tune the original CLBP algorithm to obtain rotation invariance.To achieve the robustness to scale variations,the sequential Gaussian filtering is applied to the texture image to construct a multiple-scale space,and a nonlinear operation is implemented by taking maximum values of joint histograms across different scales.Moreover,the patterns corresponding to multiple radii are combined to acquire macro and micro texture features.Experimental results show that the proposed method can extract discriminative texture features that achieve very high recognition accuracy on many benchmark texture databases,and outperform many state-of-the-art feature extraction methods of texture images.Meanwhile,the proposed method is efficient and robust to illumination,rotation,the scale variations and the number of training samples.Therefore,the proposed method exhibits excellent comprehensive performance in the feature extraction of texture images.(4)A method based on the fusion of coarse color feature and texture feature is proposed for feature extraction of color texture images.To extract color feature and maintain high efficiency,the H and S components in HSV color space are coarsely quantized due to low frequency and regional distribution of color information.The joint histogram of quantized H and S components is used as color description,and texture feature is extracted from the V component.Finally,the extracted color and texture features are combined to describe the features of color texture images.Experimental results show that the proposed method can extract the color texture features that are more discriminative than pure texture feature or pure color feature.Therefore,the proposed method can further improve the recognition accuracy,and maintain high efficiency.In addition,the proposed method achieves good performance in the applications of tree bark classification,content-based image retrieval,and banknote discrimination.