Research On Subspace Face Recognition Based On Data Reconstruction

Face recognition system consists of face detection, image preprocessing, feature extraction and classification, of which feature extraction is the most important part. Subspace method stands out from the feature extraction methods. Subspace method aims to find the best projections according to some specific performance requirements, which are always related to the way of data reconstruction. Different data reconstruction methods also lead to different subspaces. In this paper, studies mainly focus on data reconstruction and subspace methods, which takes a deep research and analysis on the advantages and disadvantages of subspace methods. Our main work includes the two aspects:(1) Sparsity Preserving Projections(SPP) reconstructs data using sparse representation, which preserves the sparse structure in the low dimensional space. However, SPP doesn’t make the most of label information of samples or distinguish the samples from different classes. A novel method called Sparse Reconstructive Discriminant Analysis(SRDA) is proposed in this paper to compensate for these shortages. SRDA takes full use of the label information, which minimizes the intra-class sparse reconstruction error. It also takes the local inter-class samples into account, simultaneously maximizing the neighbor inter-class sparse reconstruction error. SRDA improves the discriminant performance and separability between different classes. Experimental results on three large scale face databases(Yale B, PIE and AR) demonstrate the effectiveness of the proposed method.(2) Reconstructive Discriminant Analysis(RDA) reconstructs data by using all other samples from the same class. It minimizes the intra-class reconstruction scatter as well as maximizes the inter-class reconstruction scatter. However, RDA ignores the different distances between linear subspaces, resulting in poor ability to maintain data structure in subspace, and RDA is non-orthogonal, which makes it difficult to reconstruct data. A novel method called Weighted Orthogonal Reconstructive Discriminant Analysis(WORDA) is proposed to compensate for these shortages. According to the different distances between linear subspaces, the proposed method introduces weight of linear subspace to preserve better structure in subspace. Simultaneously it adds orthogonal constraint to projection to overcome the metric distortion problem, which helps get better projections in subspace. The results of extensive experiments on ORL, Yale, AR and FERET face databases demonstrate the superiority of the proposed method.