Research On Face Recognition Algorithms Based On Sparse Representation And Discriminative Dictionary Learning

In recent decades,face recognition has been an active research topic in computer vision,machine learning and artificial intelligence.The successful application of Sparse Representation-based Classification（SRC）algorithm in face recognition has attracted extensive attention from scholars to signal sparse representation.It uses a linear combination of a few number of atoms in an overcomplete dictionary containing all training samples to describe the given signal.Due to the large number and high dimension of the original training samples,the size of the dictionary obtained is large,which leads to the increase of the computational complexity of sparse coding,and the discriminative information contained in the samples cannot be fully mined.Therefore,how to learn an appropriate dictionary from the original training samples to efficiently extract the discriminative features of face data is an urgent problem to be solved in signal sparse representation.In addition,as an effective method to deal with high-dimensional data,data dimensionality reduction technology transforms high-dimensional data into a new low-dimensional subspace to remove irrelevant or redundant features,which overcomes the impact of dimensionality disaster on the performance of algorithms.Based on sparse representation and dictionary learning algorithms,this thesis further combines projection learning to study two improved face recognition algorithms.The main work contents are as follows:（1）A face recognition algorithm based on coupled discriminative dictionary learning is proposed to address the problem that the discriminative performance of analysis dictionary and synthesis dictionary is insufficient in coupled dictionary learning algorithms.In order to learn more compact and discriminative couple dictionary,the discriminative constraint term of the coding coefficients is introduced into the low-dimensional subspace to ensure its intra-class compactness and inter-class separability.Meanwhile,the graph regularization term is constructed by using the synthesis dictionary atoms to preserve the local geometric information between the data.In addition,the structural consistency term of the projection samples is introduced to constrain each class of projection samples to have the common sparse structure,so that the final recognition focuses on more discriminative low-dimensional data features.（2）Since manifold learning can accurately reveal the potential manifold structure of data,a face recognition algorithm based on sparse representation algorithm and manifold learning is proposed,which combines sparse representation and manifold learning to preserve both local and global structural information of the data.Specifically,in the original sparse representation framework,the l_2,1norm is used to replace the l₁norm constraint to represent the coefficient matrix,and then the sparsity is further used to construct the similarity matrix of the manifold regularization constraint term with low-rank embedding to select the projection samples adaptively.Meanwhile,the discriminative analysis constraint of the projection samples is introduced into the model to further improve the performance of the algorithm.In addition,in order to solve the objective function of the proposed algorithm,an effective optimization algorithm is proposed.To verify the effectiveness of the above algorithms,simulation experiments are carried out on CMU PIE,Extended Yale B,ORL and AR face databases.The experimental results show that the two proposed face recognition algorithms have achieved better recognition results compared with some related classical algorithms and recently proposed algorithms.