COC-LBP: Complete Orthogonally Combined Local Binary Pattern for Face Recognition

Abstract

Literature reports several local descriptors where discriminativity is enhanced by updating or modifying the previous techniques. This work suggests such novel descriptor (in light variations) by updating the methodology of OC-LBP. In OC-LBP, the difference among orthogonal neighbors and center pixel (of respective group) is thresholded to 1 or 0 according to thresholding function, which leads to 2 OLBP codes by putting binomial weights and adding values. Then both codes are concatenated to form OC-LBP size. The major drawback of OC-LBP is the thresholding of only Sign (S) component. This limits discriminativity of OC-LBP. To build more productive descriptor so called Complete (C) OC-LBP (COC-LBP), the Magnitude (M) details are incorporated with the S component of OC-LBP. The M based OC-LBP is called as MOC-LBP and S based is called as SOC-LBP. In MOC-LBP, the absolute of differences among orthogonal neighbors and center pixel (of respective group) is thresholded to 1 or 0 according to thresholding function. In thresholding function this time mean is utilized for comparison rather than center pixel, which is used in SOC-LBP. The two evolved MOLBP codes are concatenated to form the MOC-LBP size. Finally SOC-LBP and MOC-LBP codes are concatenated to build the COC-LBP size. Moving constantly in all positions builds the SOC-LBP and MOC-LBP images. The region feature concatenation of SOC-LBP and MOC-LBP makes COC-LBP size. The huge feature size is compressed by FLDA and matching is assisted from SVMs. On 2 benchmark datasets YB and EYB, COC-LBP marks its influence by defeating the results of many descriptors.