Texture Image Classification Based on Local Sorted Difference Refinement Pattern

The traditional local binary pattern and its variants have some shortcomings,such as high feature dimensions,not fully considering the difference between neighboring pixels in local areas and etc.We propose a novel local sorted difference refinement pattern(LSDRP)to overcome the above mentioned deficiencies.Firstly,the texture image is filtered with different Gaussian filters according to the sampling radius,and then the local neighborhood sampling points are sorted according to the pixel value.Secondly,the difference between neighboring pixels in the local area is integrated into the corresponding weight of the sorted binary code for the LSDRP feature generation.Finally,the high-frequency patterns in the LSDRP are selected to represent the texture image and the multi-scale LSDRP feature vector are cascaded to describe the image texture.The experimental results on Outex,CUReT and UMD texture datasets show that the proposed method is simple to calculate and robust to illumination and rotation variant with low-dimensional features。It is worth noting that the classification accuracies are 100%,99.38%,and 99.72%on the TC10,TC12_000,and TC12_001 texture datasets,respectively。