A comprehensively improved local binary pattern framework for texture classification

Local binary pattern (LBP) has been widely studied for its remarkable texture feature extraction capabilities. A large number of LBP variants have been proposed. Most of these variants still use the traditional LBP^riu2 mapping method to achieve the purpose of dimensionality reduction or improvement of classification accuracy. In fact, LBP^riu2 loses much texture information. Inspired by the idea of utilizing complementary information and multi-operator fusion, a mapping method based on triple complementary information (LBP^OUS) is proposed in this paper. LBP^OUS can further use complementary texture information to provide an alternative mapping solution for LBP variants, and it can be applied to any LBP variants that require mapping. Furthermore, classification accuracy, noise robustness, and computational complexity have always been the major problems for LBP to balance. This paper proposes a comprehensively improved local binary pattern framework (CILBPF). Firstly, the alpha-trimmed mean filter (ATMF) is improved to obtain well-rounded noise robustness. Secondly, LBP^OUS is applied to LBP variants to retain richer texture information. Finally, the zero elimination (ZE) is improved to ensure that the feature dimensions are maintained at a relatively low level. Experiments are conducted on five representative public texture datasets, Outex, CUReT, UIUC, UMD, and ALOT. The results show that the proposed CILBPF can effectively improve the classification performance of LBP variants by adopting LBP^OUS while ensuring computational complexity. Meanwhile, better handling of Gaussian noise and salt-and-pepper noise is also acquired. It is worth mentioning that the proposed framework also exhibits excellent noise robustness against mixed noise.