传统的LBP算法缺少对图元的相位分析,因此不能较好地区分由图元旋转形成的同类纹理图像.文中提出了一种融合图元旋转不变性和相位统计信息的纹理分析算法.新算法利用图元旋转不变性的等价类约简纹理特征,减小纹理旋转带来的分类误差,然后利用图元的统计相位特征进一步划分纹理图像分类集,进而解决由旋转不变性带来的欠分类问题.该算法选用均值漂移作为纹理分类工具,并采用Brodatz纹理库作测试.与传统的WTGGD、LBP等算法相比,分类效果有显著的改善.%The traditional LBP algorithm lacks the phase analysis of graphical element, so it could not better distinguish the same classified texture image formed by the rotation of graphical element.The paper proposes an improved texture classification algorithm based on LBP, combining two important properties of graphical elements: rotational invariance and statistical phase distribution.The algorithm utilizes the equivalence classes about rotational invariance to reduce the texture features, and can reduce the error caused by texture rotation.It uses the statistical phase distribution to further subdivide the texture image, and can resolve the under-classifying problem caused by extracting rotational invariance feature of graphical element.The improved method can well reserve some advantages of the traditional LBP.In the experiment, the tool of texture classification is Mean Shift, and the test texture images are from Brodatz.The experiment shows that the statistical phase can well describe the direction distribution of the graphical element in the texture image.Comparing with the algorithm which only uses the rotational invariance feature, the new one can improve the rate of correct classification to 98.1 ％, which is obviously better than the traditional WTGGD and LBP.
展开▼
