Discriminative kernel sparse representation vial2regularisation for face recognition

Abstract

A kernel-based discriminative sparse representation method (KDSR) via l2-norm regularisation for robust face recognition is proposed. Sparse representation in the original sample space is usually a linear representation which does not consider the non-linear relationship of samples. To overcome this limitation, KDSR first maps the sample into high-dimensional feature space via kernel tricks and then performs the discriminative sparse representation scheme in high-dimensional feature space. KDSR can capture the non-linear relationship of samples and contains more discriminative information of samples, which makes it have good classification performance. In addition, KDSR has a closed-form solution, which makes it computationally efficient and easy to apply in practice. Experimental results on three benchmark databases demonstrate that KDSR can achieve better recognition performance than many state-of-the-art methods.