Innovative sparse representation algorithms for robust face recognition

Abstract

In this paper, we propose two innovative and computationally efficient algorithms for robust face recognition, which extend the previous Sparse Representation-based Classification (SRC) algorithm proposed by Wright et al. (2009). The two new algorithms, which are designed for both batch and online modes, operate on matrix representation of images, as opposed to vector representation in SRC, to achieve efficiency whilst maintaining the recognition performance. We first show that, by introducing a matrix representation of images, the size of the ℓ1-norm problem in SRC is reduced fromO(whN) to O(rN), where r ≪ wh and thus higher computational efficiency can be obtained. We then show that the computational efficiency can be even enhanced with an online setting where the training images arrive incrementally by exploiting the interlacing property of eigenvalues in the inner product matrix. Finally, we demonstrate the superior computational efficiency and robust performance of the proposed algorithms in both batch and online modes, as compared with the original SRC algorithm through numerous experimental studies.