Dictionary Pair Learning For Image Classification Based On Sparse Structure

In recent years,the sparse representation of signals has gained wide attention in the field of computer vision because of its robustness to noise.In the theoretical background of compressed sensing,many sparse dictionary learning methods have their own advantages.However,some traditional sparse representation algorithms are too simple to learn a specific dictionary for image characteristics,resulting in weak discriminative ability;and complex iterative calculations are needed to solve sparse coding,making the model training inefficiently.On the basis of using the traditional solving algorithms,although the subsequent dictionary learning methods can learn a dictionary with strong discriminative power for a specific task,as the researcher's demand increases,superposition of a large number of unrelated constraints cannot properly utilize the effective information in the images,which will greatly reduce the learning efficiency instead.In order to alleviate the computational pressure brought by traditional sparse decomposition algorithms,the projective dictionary pair learning method uses a synthesis dictionary and an analysis dictionary to reconstruct the original image,while maintaining the reconstructed coefficients with a diagonal block-sparsity structure.This structure accurately and clearly describes the class information of the sample features,and eliminates the process of solving the sparse coding by traditional methods,and presents good results in the image classification task.Considering that the original dictionary pair learning only focuses on the reconstruction expression of the image and the sparsity constraint on the coefficients,the sample information is not fully utilized.Based on this idea of synthesis-analysis dictionary learning,this paper improves the classification and discriminant ability,as well as the image reconstruction and expressive ability from the perspectives of the geometry and category structure of the sample,maintaining the coefficients with the block-sparsity structure simultaneously.Then the validity and superiority of these algorithms are verified on the open image datasets.The main works are summarized as the followings:1)A dictionary pair learning algorithm based on sample geometry is proposed.According to the definition of the related concepts of the graph,it is found that the graph structure can represent the local adjacency relationship between samples,and the graph regularization constraint is added in the process of training the dictionary pair,which can effectively utilize the local information of the training data and help the classification performance.Finally,the algorithm is used to verify its effectiveness and superiority in face recognition and flower classification tasks.2)A dictionary pair learning algorithm based on sample class structure is proposed.First,for the classification task,learn a pair of class-specific dictionary for each class of samples,and keep them independent between each two samples.Secondly,in order to prevent the incoherence(independence)from over-removing the common information in the sample,learn a pair of common dictionaries for all categories,which can guarantee the ability to reconstruct and express the images;Simultaneously,maintain the incoherence between the common dictionary pair and class-specific dictionary pairs,making the information contained in each dictionary pair neither repeated nor missing.Finally,the effectiveness and superiority of this algorithm are verified in the tasks of face recognition and target classification.3)A method of combining sparse autoencoder and dictionary pair learning models is proposed.In order to reduce the time spent on dictionary learning in high-dimensional data,the original data can be compressed and reduced by sparse autoencoder before the dictionary training to improve the training speed,according to the correlation inside the data.In addition,for the lower-dimensional data,joint adjustment of the number of hidden layer neurons and the sparsity parameter of sparse autoencoder can improve the classification performance of the original dictionary learning method.Finally,th e advantages of adding a sparse autoencoder are proved from these two aspects.