Novel and robust color texture descriptors for color face recognition

We propose novel and robust color texture descriptors which are based on the relative dominance of the discriminative power of the color components in a multi-channel representation of a color model. LBP-like operators are derived by fitting linear regression models in 2D color subspaces RG, and GB of a 3D RGB color space keeping in view of the relative dominance of G component over the R, and B components. The linear regression models yield three operators LBPLRG, LBPLGB, and LBPLCGB, which are jointly referred to as local binary patterns using lines (LBPL). The features of the three operators are combined to form feature vectors. To further boost the performance, the LBPL features are combined with the existing local binary pattern of color images (LBPC). Experimental results demonstrate the superiority of the proposed operators, LBPL and LBPL + LBPC over the state-of-the-art LBP-, SRC- and under certain conditions over CNN-based approaches across various classifiers and they are found to be robust to many variations in the face images.