Differentiated representation and applications to face recognition

Abstract

Sparse representation has received much attention. Sparse representation usually first determines and uses a linear combination of all training samples that can well approximate to the test sample and then assigns the test sample to the class whose training samples obtains the minimal class-residual. In this paper, we propose an idea to make all training samples of different classes represent a test sample in a more competitive way, which is more useful to distinguish the class most similar to the test sample from the other classes. Based on this idea, we design a novel method, differentiated representation method, which uses a mathematically tractable means to make representation coefficients on a test sample generated from a class quite unsuitable for other classes. We propose a new classification procedure for the designed method. Applications on face recognition and object classification demonstrate that differentiated representation is very promising. It outperforms original sparse representation methods, collaborative representation and linear regression classification.