Feature extraction based on Low-rank affinity matrix for biological recognition

The low-rank representation (LRR) was presented recently and demonstrated its effectiveness for robust subspace segmentation. This paper presents a discriminative projection method based on Low-rank affinity matrix (LRA-DP) for robust feature extraction. The affinity matrix is designed to better preserve the underlying low-rank structure of data representation revealed by LRR. The experiments on the Yale, Extended Yale B, AR face image databases and the PolyU palmprint database showed LRA-DP is always better than or comparable to other state-of-the-art methods, which means underlying low-rank structure of data representation preserved by LRA-DP is helpful for classification problem.