Graph Embedding Based Manifold Learning For Face Recognition

Face recognition is one of the most popular research fields in the world,which involves computer vision,pattern recognition and machine learning.High-dimensional data has large of nonlinear and non-structural information which brings data processing and analysis a huge challenge.Face data is distributed on high-dimensional data,and the method of dimension reduction determines the quality of feature extraction,which directly affects the results of subsequent face detection and classification tasks.Finding an effective method of dimension reduction is an important step of face recognition.Manifold Learning is an effective way for dimension reduction.In this paper,we focus on the graph embedding based manifold learning and its application in face recognition.First,we analyze some state-of-art methods in detail and discuss their shortcomings.And then two feature extraction algorithms are proposed.The main work and innovation of this paper are as follows:(1)Orthogonal Neighborhood Preserving Projections finds the optimal embedding of low-dimensional subspace by setting up homogeneous adjacency graph and minimizing the homogeneous local reconstruction errors.However,it only uses the homogeneous information,which leads to not obvious structure of heterogeneous data.Motivated by this fact,we propose a novel method called double adjacency graphs based orthogonal neighborhood preserving projections(DAG-ONPP).By introducing homogeneous and heterogeneous neighbor adjacency graphs,the homogeneous reconstructing errors will be as small as possible and the heterogeneous reconstructing errors will be more obvious after data projected in low-dimensional subspace.The results of the experiments on the ORL,Yale,Extended YaleB and PIE databases demonstrate that the proposed method can markedly improve the classification ability of the original method and outperforms the other typical methods.(2)A regularized Margin Fisher Analysis(RMFA)algorithm is proposed.First,this method decomposes the homogeneous adjacency matrix into three subspaces,principal space,noise space and null space,then regularizes them to get new subspaces.Second,we projects original data into the regularized space and get the regularized data.Then the optimal projection is found by decomposing the heterogeneous adjacency matrix of regularized data.The proposed method can release the problem caused by the noise and the limit number of training samples in Margin Fisher Analysis algorithm.Moreover,it uses two standard eigendecompositions instead of one generation eigendecomposition,which avoids the singular matrix problem.The results of the experiments on the ORL,FERET,Extended YaleB and PIE databases demonstrate that the proposed method outperforms the other feature extraction methods.