Research Of Face Recognition Based On Dictionary Learning

With the progress of the times and economic development,face recognition plays an increasingly important role in social life,and the importance of face recognition research is mainly reflected in the following two aspects: On the one hand,face recognition as a kind of image recognition task,its theoretical research results can be easily generalized to other image recognition task.On the other hand,the face recognition technology has developed up to now has been widely used in security monitoring,identity verification,human-computer interaction and other fields,and has played a major role in social public safety and national security.The development of society puts forward higher and higher requirements on the reliability and stability of face recognition technology.However,the existing face recognition technology can not fully meet the actual needs of society.Based on this urgent need,the works of this dissertation are as follows:(1)We propose a dictionary learning model based on discriminant locality preserving criteria.This model makes the dictionary and the coding coefficients both obtain discriminative ability,and discriminant locality preserving criterion preserves the local manifold structure of the samples.This model eventually learns a structured dictionary and structured coding coefficients,and the construction error and the coding coefficient error jointly supervise the classification.(2)We propose a novel orthogonal discrimination dictionary learning model,in which the orthogonality constraints of the dictionary are added to reduce the redundant information of the dictionary,and the spatial structure of the coding coefficients is constrained in the model,so that the coding coefficients have a smaller intra-class scatter.In the model,the sparsity of coding coefficients is no longer emphasized,and the computation time of the dictionary model is greatly reduced by using the F-norm as the constraint of coding coefficients.(3)We propose a face recognition method based on Volterra kernel.As a non-linear system identification tool,Volterra kernels has also been successfully applied in face recognition.However,the traditional face recognition method based on Volterra kernels will involve the inverse operation of matrix.This dissertation presents a direct discriminant analysis method.When solving the model,we use twice diagonalization process to find the Volterra kernels function which can map face signal to the optimal feature space,and avoid the matrix inversion problem of traditional method.In the classification,a voting classification strategy is adopted,and the sub-features of each column are individually classified,and uses a voting strategy to implement parent face image classification.The experimental results on Yale A,Extended Yale B,CMU PIE,AR,and LFW face databases show that the proposed method is effectiveness and robustness against block occlusion noise.