Local Binary Pattern Based on Magnitude Ranking for Texture Classification
Yijie Luo, Jiming Sa, Yuyan Song, He Jiang, Chi Zhang · 2021
Local binary patterns (LBP) are considered among the most computationally efficient high-performance texture features. There are many LBP variants that can achieve great texture classification results. For most LBP variants, we found two shortcomings. The first is that only the sign information of local difference is considered and the magnitude information of the local difference is discarded. The second is that LBP operator adopts a fixed weight in the encoding process. Based on these two points, this paper proposes a new operator, local binary pattern based on magnitude ranking (MRLBP), and proposes a feature dimension reduction method. Combining the operator and the method, we have excellent texture classification accuracy across six common datasets, with an average of around 2 percent less than the best LBP variants. More importantly, the computational complexity of MRLBP is several times lower than that of these LBP variants.