Research On The Methods Of Structured Discriminant Analysis Dictionary And Dictionary Pair Learning

In recent years,image classification methods based on dictionary learning have become a research hotspot in the field of pattern recognition.Although the synthesis dictionary has been widely used in image classification,because it needs to use 0l or 1l norm minimization technology to solve the sample coding coefficients,the time complexity is high.As a dual viewpoint of synthesis dictionary learning,analysis dictionary learning is favored by researchers with its efficient and intuitive meaning.However,how to associate the analysis dictionary atoms with class labels and learn a discriminant analysis dictionary is still an open problem.In addition,how to make full use of the useful information in the training samples to learn a more compact and discriminatory analysis-synthesis dictionary pair is still an open problem.In order to solve these problems,this paper has made an in-depth study on the learning method of the analysis dictionary and the analysis-synthesis dictionary pair,the main work is as follows:(1)In this paper,we propose a structured discriminant analysis dictionary learning(SDADL)method to learn a structured discriminant analysis dictionary that consists of the class-specific analysis sub-dictionaries.Specifically,SDADL first introduces a classification error term based on the traditional analysis dictionary learning framework,and learns an optimal linear classifier by fully utilizing the class labels to significantly improve the classification accuracy.Then,SDADL introduces a discriminative sparse code error term makes samples from the same class have similar discriminant coding coefficients.Finally,SDADL introduces a structured discriminant term to ensure that the coding coefficient matrix of the sample under the analysis dictionary transformation has strong discriminatory block diagonal structure.In addition,an efficient iterative algorithm is also presented to solve the optimization problem of SDADL.(2)In this paper,we propose a joint projection learning and structured analysis-synthesis dictionary pair learning(PLSDPL)method.Specifically,PLSDPL integrates projection learning and structured analysissynthesis dictionary pair learning into a framework,and iteratively updates the projection matrix during the dictionary pair learning process,so that the sample data projected into the low-dimensional subspace can retain the features that are more suitable for the structured analysis-synthesis dictionary pair learning.At the same time,PLSDPL also imposes a low rank constraint on each synthesis sub-dictionary to weaken the influence of noise in the sample data,making the learned dictionary pair cleaner and more compact.In addition,an efficient iterative algorithm is also presented to solve the optimization problem of PLSDPL.(3)Extensive experiments of image classification have been performed on the datasets such as Extended Yale B,CMU PIE,AR,CLD 22,Caltech 101 and Scene 15.The experimental results show that the two dictionary learning methods proposed in this paper have higher classification accuracy in face image,object image,scene image and crop leaf disease image classification compared with some previous state-of-the-art dictionary learning methods,which shows the effectiveness of the two dictionary learning methods in this paper.