Transposed discriminative low-rank representation for face recognition

Abstract

In this paper, based on Low-rank Representation (LRR) we present a new method, Transposed Discriminative Low-Rank Representation (TDLRR), for face recognition in which both training and testing images are corrupted. By adding a discriminative term into LRR function, we obtained a low-rank matrix recovery with the increase the discriminative ability between different classes. LRR of transposed data is also applied to extract the salient features of these recovered data so as to produce effective features for classification. In addition, the test samples are also corrected by using a low-rank projection matrix between the recovery results and the original training samples. Experimental results on three popular face databases demonstrate the effectiveness and robustness of our method.