Face Recognition using Graph Extreme Learning Machine with L21-norm Regularization

Abstract

Mathematical approaches including subspace learning (SL), graph Laplacian, and L-21-norm regularization can be combined effectively with baseline ELM to produce a novel extension of ELM termed Face Recognition using Graph ELM with L-21-norm Regularization (GELML21) that can improve the robustness and compactness of ELM. In this paper, subspace learnings that exploit the geometrical structure of face data were utilized to produce discriminative features. ELM activation functions and some SL approaches have almost nonlinear characteristics that enhance features extraction performance while destroying the local consistency properties. For this reason, the graph Laplacian was used to regulate the samples in the same class to similar outputs. After that, L-21-norm algorithm with proved convergence was introduced to solve the resultant optimization problem and yield distinct hidden layer that clearly enhanced the accuracy. Our experimental results have demonstrated that GELML21 possessed excellent performance in face recognition in comparison with ELM variants and several popular state-of-the-art classification methods.