Symbiosis Local Binary Model And Its Application

Local feature extraction is a hot research topic in computer vision. The development of local features in the past two decades has been putting for-ward many computer vision tasks. Local Binary Pattern (LBP) is a simple but extremely effective texture descriptor. Since its first publication, the LBP has been applied to many computer vision areas including texture classification, face recognition and so on. However, due to its small described region and small number of patterns, the LBP’s discriminative power is greatly limited. To enhance the discriminative power of the LBP and preserve its robustness to rotation and illumination, this paper conducts the following four novel works from the aspect of co-occurrence:Firstly, we propose to use spatial co-occurrence of locally adjacent LBP-s to enhance the discriminative power of the LBP. Specifically, we propose a pairwise rotation invariant co-occurrence of local binary pattern (PRI-CoLBP) and develop a simple and novel encoding method to obtain global rotation in-variance. We fully demonstrate its effectiveness on six applications, including texture classification, material classification, flower recognition, leaf recogni-tion, food recognition and scene recognition.Secondly, we propose a multi-scale joint encoding of local binary pattern (MSJ-LBP) feature to capture the correlation between different scales. Ac-cording the quantitative analysis based on information theory, there exists high correlation between the LBPs in different scales. Previous LBP and its variants individually encode each scale, and then concatenate the features in all scales into the final descriptor. In this way, they ignore the texture correlation between different scales, thus lose much discriminative power. However, our MSJ-LBP could effectively capture such correlation information. The proposed MSJ-LBP shows superior performance on several texture and material datasets. Thirdly, we explore cross-channel texture correlation by jointly encoding multi-channel LBPs, and term the proposed feature as CCLBP. In many ap-plications, color information is effective, but previous methods always process each channels individually, and ignore the texture correlation between differ-ent channels. However, our CCLBP effectively captures such cross-channel texture correlation, and shows superior performance on several texture and ma-terial databases.Finally, we introduce a novel and effective co-occurrence of local inten-sity order (CoLIO) for face recognition. To improve the discriminative power of LIO, we propose to capture spatial co-occurrence of LIO to enhance its dis-criminative power. The proposed CoLIO shows better performance than LIO, and achieves superior performance compared with the state-of-the-art methods.