Texture Image Classification Based On Local Binary Patterns

As one of the representative texture descriptors,the local binary pattern(LBP)has been widely used in computer vision and pattern recognition for its simplicity,discriminative power,computational efficiency and robustness to illumination changes,such as texture classification,image retrieval,medical image analysis,target recognition,etc.In addition,a lot of LBP variants have also been presented in recent years based on the original version of LBP.We also focus on the extension of the basic LBP in the paper.Some new ideas were introduced to deal with the problems of the LBP variants.The main work is denoded as follows:(1)The definition of the basic LBP was discussed firstly.Then,the extensions of the basic LBP were analysed in detail from two aspects,the fusion of the edge information and the fusion of the direction information.(2)A new scheme was proposed to deal with the high dimensionality and reduce the influence of noise for the operators,CS-LBP(Center Symmetric Local Binary Pattern),D-LBP(Direction Local Binary Pattern)and vDLBP(Variance Direction Local Binary Pattern).Firstly,the local neighborhood is divided into different 4-orthogonal-neighbors and the features of all the 4-orthogonal-neighbor are combined together as region description.The new method can reduce the dimensionality greatly.Then,the local ternary pattern(LTP)was introduced to reduce the impact of image noise.Experimental results on texture database show that the proposed schemes can effectively enhance the accuracy rate of the traditional LBP variants.(3)For the basic LBP,CDP(Circular Derivative Pattern)and DLBP(Derivative Local Binary operator Pattern),they can not extract the higher order derivative information.According to such shortcoming,an optimal description operator is proposed in the paper,which could combine the high order derivative information of the vertical and circular directions.Further,the strategy of fusing neighborhood points is chosen to solve the influence of the number sampling points on the dimensionality of CDP and the influence of sparsity on sampling on DLBP.Experimental results on texture database show that the enhanced methods give higher scores than their original versions.(4)A new anti-noise method was introduced to reduce the impact of noise on the LGP(Local Gradient Pattern).Firstly,two denoised methods,the average gray value and the local ternary pattern,are presented in the paper.Secondly,the complemental code was given to reduce the dimensionality of the operator.Experimental results also show the better performance of the introduced methods.