Local multiple patterns based multiresolution gray-scale and rotation invariant texture classification

Abstract

The local binary pattern (LBP) is a robust but computationally simple approach in texture analysis. However, LBP does not have enough information to discriminate among multiple patterns due to its binary patterns only comprising of 0s and 1s. Thus, a multi-resolution gray-scale and rotation invariant texture classification based on local multiple patterns (LMP) is proposed in the paper. The LMP extends binary patterns to multiple patterns, which can preserve more structural information, and be more suitable for image analysis, including the analysis of flat image areas, in which case LBP code is often a random value and thus unsuitable. In addition, several extensions to the LMP, including multi-resolution analysis with the “uniform” LMP with rotation invariance, are presented. The “uniform” LMP can be regarded as a gray-scale and rotation invariant description in various applications and its discrete occurrence histogram over a region of image is proved to be a powerful invariant texture feature. Experimental results of multi-resolution texture classification with support vector machine show that the proposed method obtains overall better performance than other common methods in several aspects of space-resolution and gray-scale variation, rotation, and noise.