A novel binary adaptive weight GSA based feature selection for face recognition using local gradient patterns, modified census transform, and local binary patterns

摘要

The present paper proposes a novel feature selection scheme for face recognition problems, employing a new modified version of the gravitational search algorithm, a recently proposed metaheuristic optimization algorithm. The feature selection scheme, which reduces the dimensionality of the set of extracted features by choosing the features with high discriminative power, has been employed in conjunction with three contemporary feature extraction algorithms popularly employed for face recognition purposes, namely local binary pattern (LBP), modified census transform (MCT), and local gradient pattern (LGP) algorithms. The feature selectionis carried out by formulating a fitness function as a ratio of the within class distance to the between class distance and then a binary version of traditional GSA is developed for solving this problem. This binary GSA (named BGSA) is further enhanced to propose a novel binary variation of GSA with dynamic adaptive inertia weight (named BAW-GSA). Six new algorithms for face recognition are proposed hybridizing BGSA or BAW-GSA with each of LBP, MCT and LGP algorithms. In each algorithm, the classification step is carried out using backpropagation neural network. The algorithms were extensively tested for five benchmark face databases (Yale A, Yale B extended, ORL, LFW and AR) and it was conclusively proven that our proposed algorithms could comfortably outperform several competing, contemporary algorithms existing in literature and, among all algorithms considered, LGP hybridized with BAW-GSA emerged as the most superior algorithm.