Research On New Methods Of Image Classification Via Sparse Representation

Image classification is a hot topic in the fields of artificial intelligence and pattern recognition.The representation learning methods,represented by Sparse Representationbased Classification(SRC),have attracted extensive attentions and been widely used in various image classification tasks.However,the traditional sparse representation-based classification methods have the following drawbacks:1)the optimization of l1-norm constrained minimization problem is too time-consuming,which is not conducive to its application in practical problems;2)The representation coefficients obtained by solving the l2-norm constraint are too dense,the sparsity and discriminative ability may be insufficient;3)It is difficult to deal with the classification problem when the resolution of image is uncertain.The above drawbacks hinder the performance of sparse representation based classification methods in actual image classification tasks.In this thesis,we concentrate on the aspects of fast sparse representation,discriminative sparse representation and image classification with uncertain resolution,and we put forward several improved approaches.The main research results are summarized as follows:(1)A image classification method based on Sparse Augmented Discriminant Sparse Representation Classification(SA-DSRC)is proposed.The traditional SRC method that uses l1-norm constraint is time-consuming for solution.The representation coefficients obtained by Collaborative Representation-based Classification(CRC)that employs l2-norm constraint are too dense,the sparsity and discrimination ability of the representation coefficients may be insufficient.The discrimination and sparsity of the representation coefficients are beneficial for improving the classification performance.The proposed method first obtains discriminative and dense representation coefficients via Discriminative Sparse Representation Classification(DSRC)model,then obtains sparse representation coefficients via a sparse representation model with l0-norm constraint,and uses the sparse representation coefficients to augment the dense representation coefficients.Finally,the test samples are classified by using the augmented representation coefficients and labels of training sample.Experimental results on several standard databases show that the SA-DSRC is effective and computationally efficient.(2)An image classification method based on Weighted Discriminant Sparse Representation Classification(WDSRC)is developed.In this method,the local and global discriminative ability of the representation are considered at the same time.The local discriminative ability is obtained by the representation coefficients of the training samples corresponding to the local constraint of the samples,so that the training samples similar to the test sample can obtain larger representation coefficients and enhance the role of similar samples in linear representation.The global discriminative ability is obtained by minimizing the correlation of all category representation results.Minimizing the correlation of all category representation results can enhance the discriminative ability of representation results,which is conducive to image classification.The experimental results on several benchmark face databases,COIL-20 object database,Stanford 40 Actions database and Oxford Flowers 102 database indicate that the local and global discriminative ability of representation can improve the performance of image classification.(3)An image classification method based on Multi-Resolution Dictionary Collaborative Representation(MRDCR)is presented.Most of the traditional representation learning methods mainly focus on a single resolution images.When the resolution of the test image is uncertain,the performance of the algorithm will decline significantly.The proposed method first uses the resolution pyramid method to expand the training images(generate training images with different resolutions),then employs the dictionary composed of these images with different resolutions to represent the test image with uncertain resolution and classifies the test sample according to the class specific reconstruction errors of the training images at all resolutions.The proposed MRDCR method inherits the computational efficiency of traditional collaborative representation-based classification method and extends it from single resolution model to multi-resolution one.Experimental results on multiple face databases and a challenging virus database show that the proposed MRDCR method can deal with image recognition with uncertain resolution,and its performance is significantly higher than that of the recently proposed Multi-resolution Dictionary Learning(MRDL)method and some classical deep learning methods.(4)An image classification method based on Local Constraint Collaborative Representation with Multi-Resolution Dictionary(LCCR-MRD)is proposed.Based on the MRDCR method,the proposed LCCR-MRD method enhances the discriminative ability of the representation coefficient by using the locality of the data,i.e.,the representation coefficient of the corresponding training sample is constrained by the similarity between the test sample and the training sample,so the training sample similar to the test sample can obtain a larger representation coefficient,which can enhance its role in linear representation.Experimental results on multiple face databases and a challenging virus database show that the proposed LCCR-MRD method is effective.When dealing with uncertain resolution image classification,its performance is better than the recently proposed MRDL,MRDCR and some classical deep learning methods.