Discriminative common vector approach based feature selection in face recognition

Abstract

A novel feature selection algorithm is proposed, which is related to the Discriminative Common Vector Approach (DCVA) utilized as a means to reduce the computational complexity of the facial recognition problem. The recognition performance of the selected features is tested with DCVA and well known subspace methods over AR and YALE face databases. Moreover, the scheme indicates that important facial parts like eyes, eyebrows, noses, and lips must be kept for recognition purposes while eliminating the pixels in cheek, chin, and forehead areas. This additional knowledge comes out in the form of T-shaped and elliptical face masks used to specify the region of interest (ROI). Hence, besides the excellent dimensionality reduction given by the use of the DCVA technique, there is an intelligent use of the original database that provides superior results even in the presence of an occlusion as it is the case when the facial images have scarves.