基于局部排序差值细化模式的纹理图像分类

Texture Image Classification Based on Local Sorted Difference Refinement Pattern

针对传统局部二值模式及其扩展算法存在特征维度高、不能充分体现局部邻域像素间差值大小信息等问题,提出一种局部排序差值细化模式(LSDRP).首先根据采样半径大小对图像进行相应规格的高斯滤波,并按灰度值将局部邻域采样点排序;然后计算局部排序邻域内像素间的差值并将其融入排序二值编码对应位置的权值中,从而生成LSDRP特征;最后选取LSDRP特征模式中的高频模式表征图像,并级联多个半径下LSDRP特征的高频模式构成图像纹理的多尺度表示.在Outex,CUReT和UMD纹理库上的实验结果表明,所提算法计算简单且能在低维度条件下有效解决纹理分类中存在的光照、旋转变化问题;特别是在TC10,TC12_000和TC12_001纹理库上仅需120维特征即可分别达到100%,99.38%和99.72%的分类精度.

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。