Face Recognition Algorithm Research Based On Compressive Sensing

As an important biometric technology, face recognition technology is one of the most challenging and attractive fileds of computer vision. It had been widely used in the actual scenario of security monitoring, authentication, etc, but there are still many difficulties and problems exist. This thesis focuses on the detailed and profound statement of the new signal analysis method compressed sensing theory and its application on face recognition. Finally, the paper presents several improvements to the SRC algorithm: DWT_SRC algorithms, mathematical model of corruption and occlusion, improved algorithm — W_SRC algorithm. The experiments indicate that,SRC algorithm and its improvement performance more excellent than the traditional algorithm. The main research work is as follows:(1) Research on the basic theoretical framework of face recognition. Include:Deeply analysis about face image acquisition, face image preprocessing, face feature extraction, classifier design. Emphasis on demonstrate theoretical analysis and experimental results of the PCA feature extraction method and SVM classifier.(2) Introduce the compressive sensing theory. Include: Deeply analysis about its overall framework, mathematical background, three steps and common application.Complete present the various key processes from mathematical principles to applications about the choice of sparse basis or over-complete dictionary, the design and choice of observation matrix, l1-norm optimization algorithm and OMP algorithm.(3) Describe and demonstrate from the theoretical framework to simulation for sparse representatioin-based classification algorithm of face recognition. Experiments indicate that SRC algorithm has advantages over NN algorithm and SVM algorithm whether recognition rate or recognition rate, because sparse representation can select the most closely represented subset for the input test image and the observations contain optimum sparse subject information of the input face. Therefore, SRC algorithm perform a higher recognition rate and more robust.(4) Propose the DWT_SRC algorithm and W_SRC algorithm for the requirement of further improvement the recognition rate and robust when corruption and occlusion exist. The introduction of wavelet transform, it can not only add an effective step of dimensionality reduction but also retain the class information of image for theinsensitivity of illumination, expressions and gestures, even increase the number of sample in a sense. For the interference of practical application, this thesis presents a mathematical model of corruption and occlusion, and its improvements. At the same time, the thesis makes the original objective function equivalent to the Lasso problem.Introducing weighted l1-norm, so that the large and small coefficients of reconstructed signal obtain the equal constraints, it makes W_SRC algorithm has a better approximation to l0-norm optimization model. Simulation results show that the DWT_SRC algorithm has a higher recognition rate and faster recognition speed than the SRC algorithm, meanwhile, the algorithm benefits face recognition in small sample case.Then, the improved noise model is more realistic and robust. Finally, the W_SRC algorithm proposed by this thesis can solve the problem of corruption and occlusion to a certain extent, the algorithm is more robust than SRC algorithm.