Research On Texture Classification Based On Improved Local Binary Patterns

Texture is a visual feature on the surface of objects in nature.Through the analysis of texture,people can distinguish the attributes and categories of things.Therefor,the research about texture features is particularly important.Local binary pattern?LBP?is widely used because of its simple and efficient about the texture feature extraction,but the accuracy of traditional local binary pattern in texture classification is too low and its robustness to noise is also very poor.So,the local binary pattern is deeply researched from two aspects: the robustness to noise and the feature extraction performance of the algorithm itself.The main works are as follows:The traditional rotation invariant uniform local binary pattern ignores some features of the non-uniform pattern which can represent important information of the image,which results in the decline of classification accuracy.To solve this problem,a multi-feature local binary pattern?MFLBP?is proposed.Secondly,based on MFLBP algorithm,the MFLBP features of different radius and neighborhood sampling points are cascaded.So,multi-scale multi-feature local binary pattern?MSFLBP?is proposed.The traditional central symmetric local binary pattern?CS-LBP?does not consider the relationship between the central pixel and the neighborhood pixel,and the missing part of the image pixel information leads to the decline of classification accuracy.To solve this problem,an enhanced central symmetric local binary pattern?ECS-LBP?is proposed.The algorithm compares the central symmetric neighborhood pixels with the central pixels separately.Only when the gray value of one pixel is greater than or equal to the gray value and the other pixel is smaller than the central pixel,the pixel is set to 1.At the same time,in order to enhance the anti-noise performance of the algorithm,the threshold T is added on the basis of ECS-LBP algorithm,and CS-LBP and ECS-LBP feature are cascaded.So,enhanced noise resistance central symmetric local binary mode?ENRCS-LBP?is proposed.By analyzing the classification results on TC00010,TC12₀00 and TC12₀01datasets,it is found that the classification accuracy of MFLBP is 1.5%?2.6%?1.94%higher than the traditional rotation invariant uniform local binary pattern.The classification accuracy of MSFLBP is 11.59%?22.11%?23.63% higher than the MFLBP algorithm,the classification accuracy of ENRCS-LBP is 9.04%?10.87%?11.84% higher than the traditional LBP.respectively.The results show that the improved MFLBP ?MSFLBP and ENRCS-LBP algorithm have stronger ability to represent image features,and the features extracted can achieve higher classification accuracy in classification.The noise resistance performance of ENRCS-LBP is better than LBP and CS-LBP by adding different degrees of salt and pepper noise and Gaussian noise into the datasets.