Research Of Sparse Represention Based Face Recognition Algorithm

Face recognition is a popular subject in the field of computer technology. It exploits biological characteristics to identify the different persons. Due to its advantages, face recognition has been widely used in computer vision, pattern recognition, image processing, multimedia processing, psychology and other fields. So it plays an extremely important role in the field of artificial intelligence.Face recognition system generally includes two main modules: face detection and face recognition, and the work of this thesis mainly focuses on face feature extraction and face recognition. Based on the research of these two parts, a NMFs+MSRC based face recognition system is put forward. Non-negative Matrix Factorization with Sparseness (NMFs) was used to control the sparsity of the input face samples, and then to build an over-completed dictionary. Mahalanobis distance was introduced in Sparse Representation based method for face Classification algorithm, which called as MSRC in this thesis, to improve the recognition efficiency for similar samples. The experimental results show that the proposed algorithm can reduce the computing complexity and reduce recognition error rates meanwhile.Firstly, this thesis gave out a review on research background and developing roadmap of face recognition. And the main kinds of face recognition algorithms were introduced and analyzed, some conclusions were also presented.A deep research was made on face feature extraction for face recognition application. Feature extraction algorithm mainly includes: principal component analysis, independent components analysis, Non-negative Matrix Factorization with Sparseness Constraints and so on. NMFs algorithm was analyzed in detail, and exploited into the suggested face recognition system.Mahalanobis distance based sparse classification algorithm was proposed in this thesis. This classification algorithm theory is based on the basic principle of compressed sensing. By constructing an over-completed dictionary, the most sparse linear representation (optimal solution vectors) of the testing samples was worked out, and the face recognition was achieved ultimately.A large number of numerical experiments were performed and analyzed in this thesis. The designed experiments are used to: verify the effectiveness of NMFs, compare MSRC with other classification algorithm, and prove its validity, compare with other face recognition algorithm. The results show that the proposed algorithm has better performance in the feature extraction and recognition. Recognition rate of the proposed NMFs+MSRC can be 97%, which was higher than normal SRC's (93%).Finally, this thesis summarized the research work, and pointed out the main direction of the future work.