Face Recognition Using Robust Manifold Learning

Face recognition has been one of the active research topics in computer vision, pattern recognition due to its wide range of applications such as in commercial and law enforcement applications. Previous works have demonstrated that face recognition performance can be improved significantly by suitable feature extraction approaches. Among the most approaches, subspace analysis techniques have become an active research area owing to its low computational cost, strong describable ability and good separability. This dissertation studies the intrinsic geometrical representation of image space by manifold learning and graph theory. The main work and contributions are as follows:First, Local Similarity and Diversity Preserving Projection (LSDP) is proposed. LSDP constructs two adjacency graphs, namely similarity graph and diversity graph, to model the geometrical structure and diversity of images, which well characterizes the modes of similarity and variability. A concise feature extraction criterion is raised by minimizing the scatter, which efficiently preserves the locality, and simultaneously maximizing the variation of images. Different from the most existing manifold learn methods, LSDP not only preserves both the geometry and diversity information of images, but also avoids over-fitting. Extensive experiments show the efficiency of the proposed method.Second, Orthogonal Local Discriminant and Information Embedding (OLDIE) is proposed. OLDIE constructs two adjacency graphs to model the local intrinsic geometry of data, one adjacency graph characterizes the local information, which includes similarity and variation of the data points belonging to the same class; another graph characterizes the margin among nearby data points belonging to different classes. A concise feature extraction is built by incorporating the local information and margin into the objective function of linear dimensionality reduction. In order to further effectively detect the intrinsic geometry of data, an orthogonal algorithm is proposed to solve the generalized eigenvalue problem. Extensive experiments show the efficiency of the proposed method.