For the merits of unconstrained, concealed, as a kind of biometric identification, face recognition has becamed the most important identity certification technology. A face recognition system is made of four steps: face detection, preprocessing, feature extraction and classification. As the core of face recognition system, feature extraction will directly affect the recognition accurate. This paper mainly research feature extraction algorithm. Summarized as follows:1. Expound the principal component analysis(PCA) and singular value decomposition(SVD), After the PCA preprocessing, date is used in locality preserving projections(LPP), which solving the singular value problem. But LPP exists the parameter selection problem in computing weight matrix and the singularity problem in solution of generalized characteristic equation. In this paper, to solve these problems, a novel algorithm called based on SVD's Parameter-less Supervised Locality Preserving Projection(PSLPP) is researched. Firstly, to solve the parameter selection problem, we use the class information of samples and the cosine distance which is more robust to outliner to describe the similarity of sample points, which make up the weight boundary, constructing parameter-less nearest neighbor graph. Secondly, to solve the singularity problem, we make the secondary projection on sample matrix by using the SVD method. Experiments based on both ORL and Yale face database demonstrate that our method is effective and has a higher recognition rate compared with locality preserving projections and parameter-less locality preserving projections.2. To solve linear inseparability problem, PCA and LPP are extended kernel principal component analysis(KPCA) and kernel locality preserving projections(KLPP) by using kernel skill. In this paper, to solve the parameter selection problem in computing weight matrix in KLPP, a novel method named parameter-less supervised kernel locality preserving projection algorithm is researched. By changing the euclidean distance to the cosine distance, which is more robust to outliner, constructing a parameter-less nearest neighbor graph, the nonlinear kernel mapping is used to map the face data into an implicit feature space,and then a linear transformation is preformed to preserve locality geometric structures of the face image,which solve the difficulty of parameter selection in computing weight matrix. Experiments based on both ORL and Yale face database demonstrate that our method is effective and has a higher recognition rate compared with the principal component analysis, kernel principal component analysis, supervised parameter-less locality preserving projections and kernel locality preserving projections. |