Research On Image Classification Algorithms Via Sparse Representation

Representation-Based Classification Methods(RBCM)is one of the hot topics in the fields of computer vision and pattern recognition,and has achieved excellent results in practical applications such as face recognition,character recognition and hyperspec-tral image classification.RBCM remains a popular approach in the domain of image classification.During the past decade,researchers have proposed a large number of representation and classification methods and applied them in different areas.Recently,with the de-velopment of related fields,improved representation based classification algorithms have achieved impressive results in image recognition.However,these methods still have some shortcomings that need to be addressed,such as how to exploit the structural information,diversity and label information of the training samples to extract richer and discriminative features,so that a simple classifier can be used in the classification stage to obtain better recognition performance;and how to enhance the sparsity of representation coefficients to improve the recognition accuracy.To accommodate the above considerations,we propose several improved representation based classification algorithms.The main achievements of this dissertation can be summarized as follows:(1)A sparsity augmented weighted collaborative representation-based classification(SA-WCRC)algorithm is proposed.In collaborative representation-based classification(CRC)method,scholars point out that it is the collaborative representation mech-anism rather than the l1-norm sparsity that makes the sparse representation-based classification(SRC)method successful for classification tasks.However,recent studies reveal that the sparsity of the representation coefficient does play a critical role in the accurate classification of test samples,thus it should not be totally overlooked due to relatively high computational cost.Moreover,the local structure of data contains more meaningful information than the global structure in some cases.Inspired by these findings,we propose the SA-WCRC method.Firstly,the representation coeffi-cients of the test sample are obtained via weighted collaborative representation and sparse representation,respectively.Secondly,we augment the coefficient obtained by weighted collaborative representation with the sparse representation.Finally,the test sample is classified according to the augmented coefficient and the label matrix of the training samples.Both the augmented coefficient and classification scheme make SA-WCRC efficient for classification.Experiments on various datasets demonstrate the superiority of our proposed SA-WCRC.(2)A sparse and collaborative-competitive representation based classification(SCCRC)approach is presented.Since the collaborative-competitive representation based clas-sification(CCRC)algorithm is obtained by introducing the competitive representation term into the CRC model,and the regularization term can promote the training sam-ples in different classes to competitively reconstruct the test sample,thus the sparsity of the coefficient in CCRC is enhanced to some extent.Nevertheless,the representa-tion coefficient of CCRC is still not sparse enough,which hampers the classification performance of CCRC.For RBCM,the sparsity of the representation coefficient has an important impact on the correct classification of test samples.To this end,our proposed SCCRC multiplies the representation coefficients calculated by SRC and CCRC,and the multiplicative fusion strategy can obtain sparser coefficient,which is more conducive to classification.Experimental results in this dissertation show that the fusion strategy can effectively improve the classification performance.Moreover,when the test samples are corrupted,SCCRC can still achieve better classification results.(3)An efficient structured dictionary learning(ESDL)algorithm is developed.In dictio-nary learning,the ideal coding matrix of training data should have block diagonal structure.Towards this end,an ideal representation coefficient matrix is introduced into the ESDL algorithm,through which the dictionary atoms are associated with the label information of the training samples,and the discrimination ability of the representation coefficient matrix is improved.Meanwhile,ESDL takes full advantage of the diversity of training samples by generating alternative training samples.In addition,unlike traditional dictionary learning algorithms,we impose the l2-norm constraint on the coding coefficients,making the optimization process of ESDL more efficient.Experimental results on benchmark datasets validate the efficacy of our proposed ESDL algorithm.