Research And Application On Label Consistency Dictionary Learning With Slack And Local Constraint For Image Classification

Dictionary learning aims to generate description information that effectively reflects the nature of various features such as color and texture of images.It has always been a key issue in image classification and recognition research.In recent years,the Label Consistent Dictionary Learning(LCDL)algorithm has attracted the attention of researchers because the embedding of label information can effectively improve the discrimination of the dictionary to the image.However,the representation coefficient of most LCDL algorithms adopts a strict “0-1”structure,so that similar images have the same feature representation,but ignore the unique information of the image itself,resulting in inaccurate image representation and poor classification effect.The slack matrix can make the strict binary constraint have a wider degree of freedom and expand the selection range of image features.However,with the increase of dictionary size,the intra-class similarity of images will be weakened while affecting the training efficiency of the algorithm.In the process of feature coding,if the local constraint mechanism can be added,the neighborhood relationship of samples in each class can be effectively maintained.Therefore,in the face of image classification tasks,in order to effectively improve the discriminative ability and training efficiency of the dictionary,this paper deeply studies the label consistency dictionary learning algorithm based on relaxation and constraint.The main research work is as follows :(1)Aiming at the problem that the insufficient utilization of the unique feature information of similar images in the label consistency dictionary learning algorithm leads to the low discrimination of the representation coefficient and the poor classification effect,an image classification algorithm based on Slack Label Consistent Dictionary Pair Learning(SLC-DPL)is proposed.Firstly,the projection dictionary is used to linearly project the sample data,so that the coding representation coefficient can obtain more effective coding features while maintaining sparsity.Secondly,the label consistency matrix is introduced to associate the label information with each dictionary term.Finally,a slack matrix is learned to dynamically optimize the label consistency,so that the learned representation coefficient is more flexible,thereby improving the discriminability of the dictionary pair algorithm.Experiments on AR dataset,CMU PIE dataset and USPS dataset show that the algorithm can effectively improve the accuracy of image classification.(2)Although the SLC-DPL algorithm proposed in research(1)makes full use of the structural information of sparse coding coefficient to improve the classification effect,the dictionary pair learning method is used to make the dictionary scale larger.When the dictionary scale increases to a certain extent,the training phase will take a long time and the intra-class similarity of the image decreases.In order to further improve the accuracy and efficiency of image classification,an image classification algorithm based on Local Constraint and Slack Label Consistency Transform Learning(SLCTL)is proposed.Firstly,on the basis of slack label consistency dictionary learning,the transform matrix is used to represent the sample data to quickly capture more effective features of the sample.Secondly,an adaptive weight is learned to constrain the representation coefficient locally,which improves the similarity of intra-class representation and makes the learned dictionary more discriminative.Finally,the algorithm is verified on the USPS dataset,Scene15 dataset and CMU PIE dataset.The experimental results show that the proposed algorithm is superior to the SLC-DPL algorithm in image classification accuracy and training efficiency.(3)Based on the research methods of(1)and(2),an image classification system based on label consistency dictionary learning is designed and implemented on the Matlab platform.