A Completed Multi-Scale Local Statistics Pattern for Texture Classification

Binary pattern methods play a vital role in extracting texture features. However, most of existing methods struggle to capture comprehensive and discriminative texture information. This paper aims to propose a novel multi-statistic binary pattern to extract rotation invariance statistic features for texture classification. First, this paper encodes the center pixel, mean, variance and range of local neighborhood by corresponding multi-scale threshold, and proposes the local center pattern, local mean pattern, local variance pattern and local range pattern. Then, based on the compact multi-pattern encoding strategy, the four sub-patterns are jointly encoded in a 4-bit binary pattern, named as multi-scale local statistics pattern. Finally, for comprehensive texture representation, the multi-scale local statistics pattern is jointly combined with local sign pattern and local magnitude pattern to generate a completed multi-scale local statistics pattern for texture classification. Extensive experiments conducted on three representative databases demonstrate that the proposed completed multi-scale local statistics pattern achieves competitive classification performance compared with other state-of-the-art approaches.