In recent years,with the advancement of image acquisition technology and the popularization of video surveillance equipment,it is convenient to collect multiple images of a person to form an image set.In practical applications,there are various interference factors such as illumination and posture in the collected images,so it is difficult for the classification method based on a single image to accurately classify the image set.In addition,only a frontal image or an image set of the target is prestored in the identity authentication system,and the lack of intra-class samples further increases the difficulty of image set classification.In view of the above problems,this paper studies the classification of undersampled image sets based on sparse representation technology,and improves from two aspects of dictionary construction and dimensionality reduction,aiming to improve the accuracy of image set recognition.The specific work is as follows:(1)In Sparse Representation based Classification(SRC),the basic dictionary lacks discriminativeness and the information in the auxiliary dictionary is insufficient.For these problems,this paper proposes a single-sample face recognition method based on Kernel Extended Hybrid Block Dictionary(KEHBD).Firstly,the image is divided into blocks,and the discriminative features of each block image of the standard sample of the object to be recognized are extracted by the kernel discriminant analysis algorithm to construct the basic block dictionary;Then,the KDA algorithm is used to extract the occlusion information of the non-identified objects and the intra-class variation information such as expression and illumination,and construct the occlusion block dictionary and the intra-class difference block dictionary,which effectively alleviates the problem of insufficient intraclass sample data in the object to be recognized.Finally,the basic block dictionary,the occlusion block dictionary and the intra-class difference block dictionary are jointly used for sparse representation classification,the residual sum of all blocks is calculated,and the categories are divided according to the smallest residual.The experimental results show that KEHBD has better performance in both single image classification and image set classification.(2)Traditional image set classification methods usually use kernel function to map the complex nonlinear Riemannian manifold data to reproducing kernel Hilbert space and perform discriminative dimension reduction,which will lead to high computational complexity of the algorithm.It is difficult to choose the kernel function that satisfy the distribution of Riemannian manifold data.To solve these problems,this paper proposes an image set dimensionality reduction algorithm based on Robust Sparse Preserving Discriminant Analysis(RSPDA).Firstly,Gaussian Mixture Models(GMM)is used to model the original image set to obtain multiple SPD matrix.Then,Log operation is used to map the SPD matrix into the tangent space.Class labels and intra-class compactness constraints are introduced to optimize the sparse representation coefficient of each sample in the tangent space,and thus build a sparse adjacency graph.Finally,according to the above sparse adjacency graph,optimize and solve the best projection matrix on the basis of maintaining the distribution characteristics of the data tangent space to realize low-dimensional projection of high-dimensional image set data.A large number of experimental results verify the effectiveness of the RSPDA algorithm.(3)In view of the effectiveness of the auxiliary dictionary in the above KEHBD algorithm and the advantages of the RSPDA algorithm for dimensionality reduction of image set data,this paper proposes the image set classification algorithm based on Joint Robust Sparse Preserving and Auxiliary Dictionary Learning(JSPADL).First,an auxiliary dictionary with intra-class difference information is introduced into the construction process of the sparse adjacency graph of the RSPDA algorithm,and the sparse representation coefficient and auxiliary dictionary of the query sample are jointly optimized to make the constructed sparse adjacency graph more compact to further improve the discrimination of image set samples in low-dimensional projection space.Secondly,according to KEHBD,the auxiliary dictionary obtained by the above optimization is introduced into the sparse representation classification of image set,which can effectively solve the problem of undersampled classification problem of image sets.Experimental results show that JSPADL is more effective than RSPDA algorithm. |