Research Of Face Recognition Algorithm Based On Independent Component Analysis Modle

Within an informationized and intellectualized era, more and more attention was paid to information security. As a friendly and accurate biometric recognition technology, face recognition becomes a great interest of researchers. This paper focuses on subspace methods and independent component analysis(ICA) is the main emphasis. Comparing to traditional principal component analysis(PCA), ICA is more capable of revealing the nature of relationship hiding in data and more suitable for extracting features.Many existing algorithms based on subspace theories conduct face representations by finding a group of basis images. PCA, which relies on second-order-statistics for finding transforming matrix, is just such a typical method. ICA as a generalization of PCA, can uncover important information hiding in the high-order statistics. The two architectures of ICA model extract local and global features in face images respectively, however any single kind features cannot represent face images perfectly. Research work in this paper mainly focuses on the following aspects:1. Feature fusion method is proposed which is based on ICA model, including serial and parallel fused features. There are two levels that are base images and representation coefficients. Via analyzing the results on three standard face databases, feasibility and effectiveness of fusion algorithm above is verified. On the whole, if input data is non-Gaussian, fusion of base images has better performance than fusion of representation coefficients.2. In this paper, the integration issue of features and classifiers is studied. From recognition rate and runtime, we compare SVM classifier with KNN classifier and Mahalanobis distance. Combined with results, we analyze advantages and disadvantages of each classifier. Through the results on ORL and YALEB database, feature fusion of base images combining with SVM classifier achieves the best recognition rate in this paper.