The application of face recognition is no longer limited in security, authentication and other fields, especially the Internet banking and smart city has provided a huge potential market for biometric recognition such as face recognition in recent years. Formed by two parts mainly as to face recognition system:face detection and feature extraction. Feature extraction plays a key role in face recognition system. The main purpose of the feature extraction is to reduce the dimension of sample set, how to extract from the face image and describe individual characteristics effectively is a key issue of face recognition research.We introduces the representative feature extraction algorithms, including subspace linear dimension reduction methods such as principal component analysis (PCA) and linear discriminant analysis (LDA),and the linear methods in manifold learning such as locality preserving projections (LPP). Based on the analysis of the advantages and drawbacks of the two kinds of algorithms, we introduce the main idea of semi-supervised learning in reference to practical application. We also explained the basic idea and steps of SDA. A new semi-supervised discriminant analysis algorithm based on manifold regularization is proposed on the basis of forementioned theories. A nearest neighbor graph was constructed first to estimate the intrinsic geometrical structure of the sample, and then the graph structure was incorporated into the objective function of the multivariate linear regression as a regularization term. At last, a set of optimal projection is obtained by maximizing the objective function. For the convenience of test, we complete the face test system based on MATLAB simulation, and introduces the system design and implementation of several modules.This main contribution of the article is presented as follows:First, The method we proposed avoid the overlap of classes which have a small class separation distance by restructuring the with-class scatter and between-class scatter matrix. Secondly, we considered the local and nonlocal geometric structure of samples simultaneously in the construction of neighbor graph, and the small sample problem can be solved due to the addition of the regular item. It is indicated that the algorithm have better robustness in the case of lacking sample data. Experimental results on ORL and Yale face recognition demonstrate the effectiveness of the algorithm. |