Sparse-representation based classification (SRC) has been showing a good performance for face recognition in recent years. But SRC is not good at face recognition with low quality images (e.g., disguised, corrupted, occluded, and so on) which often appear in practical applications. To solve the problem, in this paper, we propose a novel SRC based method for face recognition with low quality images named sparse low-rank component based representation (SLCR). In SLCR, we utilize low-rank matrix recovery on the training dataset to obtain low-rank components and non-low-rank components, which are used to construct the dictionary. The new dictionary is capable of describing facial feature better, especially for low quality face samples. Furthermore, the minimum class-wise reconstruction residual is used as the recognition rule, leading to a substantial improvement on the proposed SLCR’s performance. Extensive experiments on benchmark face databases demonstrate that the proposed method is consistently superior to other sparse-representation based approaches for face recognition with low quality images.

INTRODUCTION

F ACE recognition has been the most popular biometric method due to its huge application potential in the past decades [1], [2], [3], [4], [5], [6], [7], [8]. Sufficient and favourable training samples guarantee a good feature representation for describing the characteristics of an individual’s face. However, in the real world, the image of each person is often disguised, corrupted or occluded. Therefore, face recognition with low quality images is more challenging than the one with sufficient and favourable images. This paper focuses on the task of face recognition with low quality images. The effectiveness of feature extraction is important for face recognition. Principal component analysis (PCA) [9] is a common technique for dimensionality reduction. In addition, there are other methods such as linear discriminant analysis (LDA) [10], probabilistic subspace learning [11] and locality preservation (Laplacianface) [12] and so on. However, it is a difficult task for these methods to solve outliers or sparse noise [13]. To alleviate this problem, some methods on robust PCA have been proposed [14], [15], [16]. Among them, low-rank matrix recovery (LR) [14] is a key technique, which can separate corrupted information from the training face images better than PCA. Accordingly, lowrank components obtained by LR would better serve the classification purpose. The performance of classifier is important for face recognition. Nearest neighbor (NN) classifier is widely applied for its simplicity. Extensions of NN classifier, nearest feature line (NFL) [17], nearest feature plane (NFP) [18], nearest feature space (NFS) [19] and linear regression classifier (LRC) [20], consider the relation between the testing image and the training images of each class separately [21]. Different from the above-mentioned classifiers, sparse-representation based classification (SRC) which considers the testing image as a linear combination of the training dataset has been proposed for face recognition and achieved satisfying results [22]. However, SRC is incapable of performing well when the training dataset is undersampled or corrupted. To overcome this shortcoming, some extended SRC methods have been proposed [25], [26]. Zhou et al. applied SRC with Markov random fields to address the disguise face recognition problem with large contiguous occlusion [23]. Wagner et al. used SRC to handle the misalignement, pose and illumination invariant recognition problem [24]. Yang et al. borrowed the idea of robust regression [27] and proposed a regularized robust coding (RSC) [28], [29]. He et al. made use of the correntropy induced robust error metric and presented the correntropy based sparse-representation algorithm (CESR) [30], [31]. Lai et al. applied a method of classwise sparse-representation (CSR) to tackle the problems of the conventional sample-wise sparse-representation [33].