Face Recognition Based On Compressed Sensing Theory

Face recognition is a potentially promising biometrics technology, which extracts information from facial images to achieve identification. Face recognition has wide application in many fields such as security monitor and ID authentication, as direct access to the participants is not required, and human face is almost impossible to counterfeit, especially comparing with passwords and bar codes. On the other hand, research of face recognition has great theoretical value, involving subjects of image processing, computer vision, pattern recognition, machine learning and so on. Research on face recognition mainly focuses on feature extraction and classification.Compressive Sensing theory is a newly proposed signal codec theory. The theory indicated that as long as a signal is sparse or compressible, it can be sampled at a frequency much lower than the Nyquist frequency, and perfectly reconstructed with overwhelming probability. It is one of the research hotspots recently, and it is quite prospective in research of face recognition. Meanwhile, manifold learning is also one of the central issues in data mining as well as a promising method as feature extraction, which intends to discover the low dimensional structure lying in high dimensional data.This paper studies firstly the compressive sensing theory and the Recurrent Neural Network for solving sparse representation problems along with its improved model. Numerical experiment is conducted to investigate the influence of dimensional parameters, and also the performance of several reconstruction methods. Then sparse representation is applied to face recognition as Sparse Representation Classifier, and combined with Gabor feature and subspace methods, achieving higher results than traditional methods. On the other side, this paper studied manifold learning theory, and improved Marginal Fisher Analysis by whitening the intra-class intrinsic matrix and selecting eigenvalues to diminish over-fitting, namely, Enhanced Marginal Fisher Model. Thereafter experiments are conducted on ORL, AR and FERET database, and proved the feasibility and effectiveness of the method and framework proposed in this paper.