A Study On Face Recognition Based On Sparse Representation And Matrix Recovery

This dissertation is a study on face recognition algorithm based on sparse representation and matrix recovery,improving the face recognition based on sparse representation classification and adaptive sparse representation classification in combination with dictionary learning and low rank matrix recovery,respectively.The main contents are as follows:1.We propose a robust face recognition algorithm based on dictionary decomposition and sparse representation,which aims to improve the recognition rate of face recognition algorithm and the robustness to noise,pollution and occlusion.As the SRC obtains poor performance when there are noise,and corruptions and occlusions in the training samples or the number of training samples is insufficient,we utilize the dictionary learning theory to design the dictionary decomposition model and extract the class-specific information from the original face images to reduce the impact of pollution.The optimization method is used to solve the model and the class-specific dictionary is obtained,which is used as the representation dictionary in the SRC.We assume that the class-specific dictionary exists in a certain subspace of the original training data,then a projection matrix is learned to represent the mapping relationship between the class-specific dictionary and the original training sample,so we can correct the test samples by projecting them to the corresponding subspace.Then the eigenface method is used to reduce the dimensions of the class-specific dictionary and the corrected test samples,and finally the SRC is used to classify.The experimental results show that the proposed algorithm can achieve good performance and is robust to noise,corruptions and occlusions.2.We propose a robust face recognition algorithm based on discriminant low rank matrix recovery and adaptive sparse representation,which aims to improve the recognition rate by using the low rank matrix recovery technique to reduce the influence of pollution and disguise on face recognition.The adaptive sparse representation classification(ASRC)considers both the effect of sparseness and the effect of cooperativity,and achieves better performance than SRC.Since samples in face recognition are usually corrupted,this will affect the performance of face recognition.Inspired by the the maximum inter-class distance and minimum inner-class distance in the linear discriminant analysis,firstly,we add a similar discriminant constraint term to the traditional low rank matrix recovery model,then the discriminant low rank matrix recovery model is obtained.The optimization method is used to solve the model and the low rank matrix recovery result is obtained,which is used as the representation dictionary in the ASRC.Then a projection matrix is learned to represent the mapping relationship between the matrix recovery result and the original training samples,so we can correct the test samples by projecting them to the corresponding subspace.Then the eigenface method is used to reduce the dimensions of the matrix recovery result and the corrected test samples,and finally the ASRC is used to classify.The experimental results show that the proposed algorithm can achieve good performance and is robust to pollution and disguise.