The Research On Discrimination Constrained Collaborative Representation For Classification

Artificial intelligence and pattern recognition have made unprecedented progress with the great improvement of computing power.As atypical method of representation-based pattern recognition,collaborative representation(CR)with its efficient and effective classification performance,has attracted much attention recently.CR represents a test sample by all training samples collaboratively without considering inter-class discrimination which is crucial for classification.Hence,in order to learning class-specific discriminative representation and further improve the classification performance of representation-based methods,our main research work focuses on the following three aspects(1)For decorrelating class-specific representation in CR,a new discriminative collaborative neighbor representation(DCNR)method is proposed for image classification.DNCR not only enhances inter-class discrimination,but also exploits the data locality.In CR,the representation coefficients can reflect the similarities between the test sample and training samples.DCNR exploits distance information between the test sample and each training sample to restrain representation coefficients that makes training samples nearest to the test sample have bigger reconstruction contribution Thus,DCNR can choose nearest sample to reconstruct the test sample discriminatively In addition,DCNR is extended to l1-norm representation fidelity-based robust DCNR(R-DCNR)for recognizing noisy image.Experiments on seven data sets demonstrate the outstanding classification performance and robustness of the proposed DCNR and R-DCNR(2)Although enhancing the inter-class discrimination in CR is very helpful for classification,the inter-class discrimination fails to emphasize the similarity between the test sample and the true class.In view of this,a weighted discriminative collaborative competitive representation(WDCCR)method is proposed.In WDCCR,the classification phase of CR is integrated into the representation phase resulting in competitive representation from each class.Meanwhile,the correlations among different classes are degraded with the aid of discrimination constraint term in WDCCR Moreover,for further improving class-specific competitive and discriminative representation,two types of weight on class-specific representation coefficients are designed,which can reflect the intrinsic relationship between the test sample and each class.In the same way,l1-norm representation fidelity-based extended method of WDCCR(R-WDCCR)is proposed for verifying the ability of recognizing noisy images Extensive experiments show the promising classification performance and anti-noise ability of the proposed WDCCR and R-WDCCR(3)Most dictionary learning methods cannot effectively learn discriminative and compact information from training data.In view of this,a discriminative multi-dictionary learning(DMDL)method is proposed.There are three kinds of dictionary in DMDL,including mean dictionary,variant dictionary and generic dictionary.The mean dictionary is obtained directly by calculating mean of each class,which reflects class-specific common feature.The class-specific variant dictionary is learned from each class that preserves compact and discriminative information of each class.The generic dictionary learns the shared component from training data of all classes.Due to mutual complementation and promotion of three kinds of dictionary,DMDL can obtain the faithful representation for given sample from the true class.Compared to the state-of-the-art dictionary learning methods and representation-based classification methods,the proposed DMDL shows excellent classification results...