Abstract
Strict '0-1' block-diagonal low-rank representation is known to extract more structured information. However, it is often overlooked that a test sample from one class may be well represented by the dictionary atoms from other classes. To alleviate this problem, we propose a robust low-rank recovery algorithm (RLRR) with a distance-measure structure (DMS) for face recognition. When representing a test sample, DMS highlights the energy of the low-rank coefficients when the distance from the corresponding dictionary atoms is small. Moreover, RLRR introduces a structure-preserving regularization term to strengthen the similarity of within-class coefficients. Besides, RLRR builds a link between training and test samples to ensure the consistency of representation. The alternative direction multipliers method (ADMM) is used to optimize the proposed RLRR algorithm. Experiments on three benchmark face databases verify the superiorty of RLRR compared with state-of-the-art algorithms.