Low-rank representation based discriminative projection for robust feature extraction

The low-rank representation (LRR) was presented recently and showed effective and robust for subspace segmentation. This paper presents a LRR-based discriminative projection method (LRR-DP) for robust feature extraction, by virtue of the underlying low-rank structure of data represesntation revealed by LRR. LRR-DP seeks a linear transformation such that in the transformed space, the between-class scatter (i.e. the distance between a sample and its between-class representation prototype) is as large as possible and simultaneously, the combination of the within-class scatter (i.e. the distance between a sample and its within-class representation prototype) and the scatter of noises is as small as possible. Our experiments were done using the Yale, Extended Yale B, AR face image databases and the PolyU palmprint database, and the results show that LRR-DP is always better than or comparable to other state-of-the-art methods.