Face recognition by optimal band selection in contourlet transform domain

Face Recognition (FR) under varying pose, illumination and expression (PIE) conditions is challenging, and extracting PIE-invariant features is an effective approach to solve this problem. For enhancing the performance of an FR system, this paper proposes a unique combination of Contourlet Transform (CT), Discrete Cosine Transform (DCT) and Binary Particle Swarm Optimization (BPSO). CT and DCT are used for efficient feature extraction and a BPSO-based feature selection algorithm searches the feature space for an optimal feature subset. The optimal band selection in multiresolution and multidirectional subspaces of CT enables extraction of PIE-invariant facial features. DCT reduces the feature-vector size, lays a ground for BPSO-based feature selection, and also aids in better classification. Experimental results show the promising performance of the proposed algorithm for enhanced face recognition on eight benchmark face databases, namely Color FERET, Pointing Head Pose and UMIST (pose); Extended Yale B and CMU-PIE (illumination); JAFFE and ORL (expression); CAS-PEAL (expression and lighting). A significant increase in the recognition rate and a substantial reduction in the number of features selected are observed as compared to other FR systems.