Research On Discriminant Analysis Dictionary And Dictionary Pair Learning Methods For Image Classification

In recent years,discriminative synthesis dictionary learning,discriminative analysis dictionary learning,and discriminative analysis-synthesis dictionary pair learning have been widely used in image classification tasks,and have achieved high classification accuracy.However,in discriminative analysis dictionary learning,it is still a challenging to learn a more compact and discriminative analysis dictionary to ensure that the coding coefficient matrix of training samples presents a more discriminative block diagonal structure.In addition,in discriminative analysis-synthesis dictionary pair learning,it is worth to make an in-depth study on how to flexibly use the sample local structure information to learn a dictionary pair with powerful representation and discrimination capabilities.To address these issues,this paper has made an in-depth study of discriminative analysis dictionary learning methods and discriminative analysis-synthesis dictionary pair learning methods.The main work is as follows:(1)In order to ensure that the coding coefficient matrix presents a more discriminative block diagonal structure and enhance the discrimination capability of the analysis dictionary,we propose a self-eliminating discriminant analysis dictionary learning(SeDADL)method.Specifically,we first design a novel analysis dictionary regularization term to improve the discrimination capability of analysis dictionary by eliminating repeated and linearly dependent atoms in the analysis dictionary while preventing the generation of trivial solutions.Then,we design a self-eliminating coding coefficient constraint term to enhance the discrimination capability of spare codes by forcing the coding coefficient matrix to achieve an approximate block diagonal structure.In order to further improve the classification efficiency of SeDADL model,we introduce a linear classification error term into SeDADL model to learn a linear classifier,which constructs the links between spare codes and class labels.Moreover,an efficient iterative algorithm is designed to solve the optimization problem of SeDADL.(2)In order to fully utilize the sample local structure information and the structured information to enhance the discrimination capability of analysis dictionary,we propose a joint structured constraint discriminant analysis dictionary learning(JSCDADL)method.Specifically,we first design an adaptive local structure preserving term to enhance the discrimination capability of analysis dictionary by adaptively transferring the sample local structure information to ensure that the similar samples have similar coding coefficients under the action of analysis dictionary.Then,we design a discriminative sparse coding error term that forces the coding coefficient matrix obtained by samples under the action of analysis dictionary to have the desired block diagonal structure.Finally,we design an analysis dictionary combined term to further enhance the discrimination capability of analysis dictionary by constantly approximating the two learned analysis dictionaries to obtain an analysis dictionary with the sample local structure information and the structured information.Moreover,an effective iterative algorithm is designed to solve the optimization problem of JSCDADL.(3)In order to flexibly utilize the sample local structure information and enhance the discrimination and representation capabilities of the analysis-synthesis dictionary pair,we propose an adaptive structured analysis-synthesis dictionary pair learning(ASDPL)method.Specifically,we first design an adaptive local structure preserving term to flexibly and accurately transfer the sample local structure information to the coding coefficients to ensure that the coding coefficients of samples from the same class are similar for enhancing the discrimination capability of spare codes and dictionary pair.Then,we introduce a discriminative sparse coding error term to ensure that the coding coefficient matrix has a block diagonal structure with powerful discrimination to enhance the discrimination of the analysis-synthesis dictionary pair indirectly.Finally,we build the ASDPL model by integrating the two constraint terms into the basic discriminative analysis-synthesis dictionary pair model.Moreover,an effective iterative algorithm is designed to solve the optimization problem of ASDPL.(4)We have performed comparison experiments on six datasets,including three face image datasets Extended Yale B,CMU PIE and AR,a crop leaf disease image dataset CLD 22,an object image dataset Caltech 101,and a scene image dataset Scene 15.The experimental results show that SeDADL,JSCDADL and ASDPL methods can achieve better classification results when dealing with image classification tasks.