Study For Image Recognition Via Adaptive Sparse Representation

In recent years,image recognition has matured due to the enhancement of image acquisition technology and data processing capabilities,and has brought more convenience to people's lives.Sparse representation-based classification is one of the representative studies.This method linearly represents test sample with training samples,and a sparse constraint is used to restrain the representation coefficients,then predicts the label of class of test sample according to the classification criterion.For sparse representation-based classification,a suitable regularization plays an important role in the constraint of the representation coefficients.The traditional constraint based on L1-norm and L2-norm ignores the category information of training samples,while the constraint base on group sparse lacks the consideration of the difference of distribution structure intra-class.Through research,it is found that the separation problem in intra-class of samples has a significant influence on the regular effect.The suitable regular form helps to obtain a robust representation coefficient.Moreover,the regularization method base on L2-norm constraint is also an important research direction in the sparse representation-based classification.The L2-norm regularization has the characteristic of fast solution speed,but the sparseness of the produced representation coefficients is not conducive to classification.In view of the above problems,this paper gives the corresponding solutions and obtains the following research results:1.This paper proposes an Adaptive Micro Elastic Net regularization method.This regularization imposes group sparse and sample sparse on representation coefficients according to the specific distribution of samples.Moreover,the adaptive parameter of group sparse and L1-norm is obtained according the specific distribution of samples.Experiments show that this method can effectively improve the recognition rate.2.This paper proposes a Sparse Representation Classification via Adaptive Micro Elastic Net Regularization method and Robust Sparse Representation Classification via Adaptive Micro Elastic Net Regularization method,in which robust classification method applies M-estimators Huber loss function as the fidelity function.Moreover,an optimization algorithm base on the Alternating Direction Method of Multipliers is proposed to solve the proposed method.3.This paper proposes a L2-norm reconstruction sample constraint sparse representation-based classification method.On the basis of retaining the original L2-norm coefficient constraint,the method appends L2-norm constraint on classspecific reconstruction samples.Experiments show that the coefficients produced by this method have better discriminability than the L2-norm constraint and are favorable for classification.