In the past few decades,face recognition has been a key research field in academia and industry,and has been successfully applied to various industries and has obtained a wide range of social benefits,such as identity recognition,video security,and face payment.Due to complex situations such as different lighting,expressions,angles,poses,and occlusions,and at the same time restricted by the insufficient number of samples,face recognition still faces some difficulties and challenges in practical applications.Sparse representation and dictionary learning is an important research direction of face recognition.Because of its excellent performance in data dimensionality reduction,feature extraction,and classification and recognition,it has become an important method for face recognition research.This article mainly focuses on the problems of different lighting,expressions,postures and occlusions,insufficient samples,and practical application scenarios.The main research contents are as follows:(1)An improved image classification algorithm is proposed.By generating virtual samples through novel image representation methods,the large-scale information and global features of the original image can be better preserved,and the difference of the same object in different images can be reduced.The classification scores corresponding to the original image and the virtual image are obtained through sparse representation,and then a simple and efficient score fusion scheme is used to fuse the two classes of classification scores to obtain the final classification score to achieve the classification of the test sample.(2)A dictionary learning algorithm based on dictionary reconstruction is proposed namely the dictionary re-learning algorithm.The algorithm is based on a dictionary discriminant model of atomic locality constraints and label embedding constraints.First,use the original training samples to initialize the dictionary and coding coefficient matrix.Then,the best dictionary and coding coefficient matrix are obtained through the iterative update,and the product of the best dictionary and the coefficient matrix is used to reconstruct the training samples.The reconstructed training samples are used to re-learn the dictionary and coding coefficient matrix and implement subsequent image classification.At present,most of the dictionary learning algorithms learn the dictionary directly from the original training samples,ignoring some noise in the training samples themselves.This dictionary reconstruction method can effectively and partially eliminate the noise in the original training samples.(3)A multi-resolution dictionary learning algorithm based on sample expansion is proposed.Considering the impact of different resolution images on the performance of the dictionary learning algorithm and the small number of training samples will reduce the performance of the dictionary learning algorithm.In this paper,the original training samples are converted into different resolutions,and a corresponding dictionary is generated for each resolution image.Similarly,the original training samples of different resolutions are generated into virtual samples of the same structure,and a corresponding dictionary is generated for the virtual samples of each resolution.Finally,a score fusion scheme is used for the classification scores corresponding to the training sample and the virtual sample to obtain the final classification distance of the test sample.The virtual samples are generated by the image representation method,which effectively expands the number of training samples and alleviates the problem of insufficient samples. |