Research On Texture Image Classification Method Based On Refinement Of Local Binary Pattern

Texture classification,as one of the main tasks of texture analysis,has received extensive attention in computer vision and image processing.In texture classification,good feature representation requires not only high degree of classification,but also robustness to various image transformations.In the past few decades,many methods for texture feature description have been proposed.Among them,the Local Binary Pattern(LBP)has been successfully applied to texture image feature representation because of its simple calculation and low complexity.However,LBP only encodes the difference sign between the center point and the adjacent sampling points in the local neighborhood of the image,so its feature representation is too simple to describe the local texture details.Therefore,from the perspective of refining LBP features,two methods of feature extraction which can effectively describe local texture details are proposed in this paper.Specifically,the research content of this paper are as follows:(1)For the problems that traditional LBP and its variants have high feature dimensions and cannot fully reflect the difference magnitude between the local central point and its adjacent sampling point,a local sorted difference refinement pattern(LSDRP)is proposed.Firstly,the local sorted neighborhood is obtained by descending order of the local neighborhood sampling points according to the pixel gray value;Secondly,the LSDRP feature is generated by calculating the difference between the center point and the adjacent sampling point in the local sorted neighborhood and incorporating the difference into the weight of the corresponding position of the sorted binary code;Finally,the dimension reduction of LSDRP,that is,the high-frequency pattern of LSDRP feature pattern is chosen to represent the texture image and used in classification experiment.Experimental results on three texture databases show that the proposed method is simple to calculate and can effectively solve the illumination and rotation problems in texture classification under lowdimensional conditions.(2)Aiming at the problem that traditional LBP and its variants cannot effectively describe and distinguish between weak local patterns and strong local patterns with the same feature label in texture images,a texture feature extraction method based on global information to refine local patterns is proposed.This method includes two descriptors: the magnitude refined local sign binary pattern(MRLBP?S)and the center refined local magnitude binary pattern(CRLBP?M).MRLBP?S uses the local difference magnitudes in the entire image to assign local neighborhoods with the same difference sign as feature labels with strong and weak contrast differences;CRLBP?M uses the local central gray value in the entire image to assign local neighborhoods with the same difference magnitude as feature labels with strong and weak gray differences;Finally,the feature histograms of MRLBP?S and CRLBP?M are cascaded to construct texture descriptor MCRLBP.The experimental results on five texture databases show that the classification performance of the proposed method is better than most texture classification methods based on LBP.(3)In order to reduce the noise,we propose to perform Gauss filtering on the texture image according to the size of local sampling radius before obtaining LSDRP and MCRLBP feature descriptors.