Research On Image Classification And Feature Extraction Algorithms Under Sparse Constraints

In recent years,sparse representation has become a research hotspot in the fields of signal processing,computer vision and pattern recognition,which not only has good theoretical development,but also has excellent practical application value.Especially in the field of pattern recognition,sparse representation can not only be used as an effective image classifier,but also can be applied to the dimension reduction of data to realize the potential choice of feature.In this manuscript,based on the theory of sparse representation,in view of image recognition,especially biometric image recognition,sparse representation classifier,feature extraction,and low-resolution image feature matching problems are studied.The main work and innovation points can be divided into the following four aspects:(1)To avoid the defect that the sparse representation classifier may produce false classification in the multiple peak coefficients,an adaptive sparse representation based classifier(ASRC)based on swarm intelligent optimization and block dictionary was proposed.In order to reduce the misjudgment,the sparse coefficient emphasis term was added in the optimization objective,which made the coefficients more concentrated in the corresponding position of the similar training samples,and the nonconvex model was solved by artificial swarm intelligence.The innovation of this work lies in:putting the coefficient emphasis item directly into the objective function,generating more outstanding peak value automatically during the optimization of the coefficients,weakening the peak of jamming;the adaptive Artificial colony algorithm and the dictionary block strategy are used to improve the efficiency of optimization calculation.(2)A group collaborative representation based classifier(GCRC)with group constraint capability was proposed to recognize the images of finger-knuckle-print.The model has the advantage of fast solution,although the coefficients were not sparse enough,but with good reconstruction.By introducing group regularization,the penalty of regular term to coefficients was different in each class.The recognition experiment on the finger-knuckle-print showed that GCRC was high discriminative and high efficient.The innovation of this work:Different kinds of samples are given to different regularization penalty regulars,so that the synergistic representation coefficients have a more obvious grouping structure.(3)The sparse constraint is applied to the dimensionality reduction algorithm of tensor data,and a tensor locally preserving sparse projection(TLPSP)was proposed.The image was considered as a two-dimensional tensor,and the sparse projection of tensor was solved under the criterion of preserving the invariant locally manifold.In theory,a simple equivalent regression form was proposed,which avoided the large size of the dictionary and simplifies the model.Good performance has been achieved in both supervised recognition and unsupervised clustering applications.The innovation of this work:This paper proposes an equivalent regression theorem in tensor mode and gives a detailed derivation to prove that different from vector mode,the proposed new regression form has a smaller regression matrix scale.Although the theorem is proposed to simplify the regression of tensor objects,it is also suitable for one-dimensional vector objects,which is a simple and universal regression model.(4)The L2,1-norm constraint was introduced into the common space coupled projection algorithm,and an L2,1-norm regularized multidimensional scaling(MDSL21)model was proposed to achieve high-low resolution feature matching with feature selection capability.By using the idea of graph embedding to find the common responses,the L2,1-norm constraint was introduced into the objective function of the projection matrix,and the sparse coupled projections with feature selection ability can be obtained,which made the low-resolution face recognition performance better.The innovation of this work:The idea of sparse projection is extended from single subspace learning to coupling subspace learning,which can be used to inspire a unified framework of coupled sparse projections,and to provide a sparse extension of the heterogeneous face recognition algorithm.