On a development of sparse PCA method for face recognition problem

Abstract

Face recognition is a very significant branch of application in the pattern recognition area. It has multiple applications in the military and finance, to name a few. In reality, sometimes we have to deal with sparse representation of the facial data. In this paper, to deal with sparsity, two advanced versions of Sparse PCA are developed, including Proximal Gradient Sparse PCA (PG Sparse PCA) and Fast Iterative Shrinkage-Thresholding Algorithm Sparse PCA (FISTA Sparse PCA). Then they will be respectively applied to solve the face recognition problem by considering their combination with nearest-neighbor method and with the kernel ridge regression method. Experimental results illustrate that the accuracies of PG Sparse PCA and FISTA Sparse PCA are equivalent in the combination with nearest-neighbor, while PG Sparse PCA performs better than FISTA Sparse PCA in the case of using kernel ridge regression. However, we recognize that the computing process of FISTA Sparse PCA, on average, is always faster than the PG sparse PCA version due to the use of a fast proximal gradient version.