Research On Face Recognition Algorithms Via Sparse Representation

In recent decades, face recognition biometric technology occupy a very important position, gaining more and more attention, has become a hot research field of computer science. The key steps of face recognition technology are feature extraction and dimensionality reduction. The researchers proposed some classic feature extraction and dimensionality reduction methods, such as principal component analysis(PCA), linear discriminant analysis(LDA) and locality preserving projections(LPP) and so on. Especially in recent years, the sparse representation in the field of pattern recognition and computer vision has attracted wide attention. Sparse representation using overcomplete dictionary to extract sparse characteristics of the sample. The classic method uses a sparse representation of all the training samples as a dictionary, this dictionary contains miscellaneous information, which not only increases the amount of computation, and a correlation between the characteristics of the different categories. In this paper, proposed the methods based on improving the sparse representation dictionary, and integration with other features of the method. The main work includes three aspects.A first aspect of the work of this paper is to propose the use of non-negative matrix factorization(NMF) method to generate a dictionary sparse representation. This dictionary is the basis matric, which is obtained by using non-negative matrix factorization on training samples of each class. This eliminates the correlation between the characteristics of different classes, and increased feature sparsity, also reduced the computational complexity. Based on this dictionary, this article proposed feature fusion on block NMF(BNMF) and the proposed non-negative sparse representation based on BNMF(NSR), namely NSR + BNMF. This method combines the features of the previous two methods, increasing the richness of features, experimental results on ORL and FERET database also verified the excellent performance of the proposed method.In the above method, the non-negative matrix factorization approximation process will lose part of the information. A second aspect of the work of this paper is to present a study based on QR decomposition dictionary sparse representation. This method not only retains the excellent performance of the above methods, while increasing the accuracy of the algorithm. And this method does not require non-negative constraints, so the calculation process is simple with a certain degree. The results also show that this method has a greater degree of performance improvement over previous methods.A third aspect of the work of this paper is to propose a direct LDA(DLDA) and block-based QR decomposition of sparse feature(ASQR) fusion of representation, namely DLDA + ASQR. DLDA is a more direct and accurate information and retain more features and algorithms category information. By adjusting the ratio of the two methods of feature fusion experiments, the best feature fusion methods is obtained. Tested on ORL and FERET database proved that this method has higher recognition than DLAD and ASQR these two methods.