Research On The Method Of Dictionary Learning Under Sparse Representation Framework

In recent years,sparse representation-based image recognition is a hot topic in the field of pattern recognition,and dictionary learning is a key issue of sparse representation-based image recognition.Researchers have proposed lots of dictionary learning methods for image recognition.Although most of the existing dictionary learning methods have achieved promising performance in image recognition,they do not perform well when the training samples are corrupted with relatively large noise.In order to solve this issue,this thesis systematically studies the dictionary learning method,and the main works are as follows:1)This thesis proposes a discriminative low-rank graph preserving dictionary learning(DLRGP_DL)method with Schatten-p quasi-norm regularization,which attempts to learn a structured dictionary that has powerful representational and discriminative ability by means of low-rank constraint transformation and local geometrical structure preserving.Specifically,this thesis imposes the Schatten-p quasi-norm regularization on sub-dictionaries to make them to be of low-rank,which can effectively reduce the negative effect of noise contained in training samples and make the learned dictionary pure and compact.To improve the discriminative capability of the learned dictionary,this thesis applies a discriminative graph preserving criterion to coding coefficients during the dictionary learning process with the goal that the similar training samples from the same class have similar coding coefficients.Given a query sample,the learned dictionary is first used for sparse coding,and then both the learned coding coefficients of training samples and the class-specific reconstruction errors are used for classification.2)This thesis proposes a low rank graph preserving discriminative structured dictionary learning(LRGPDSDL)method,which puts forward further requirements on the types and relations of sub-dictionaries.Specifically,this method learns a structured dictionary that consists a common sub-dictionary shared by all the classes and the class-specific sub-dictionaries belonging to the corresponding classes.The common sub-dictionary is used to describe the common features shared by all classes,and the class-specific sub-dictionaries are responsible for characterizing the particular features in the corresponding classes.Owing to the common sub-dictionary,the learned class-specific sub-dictionaries are more discriminative and suitable for classification tasks.In addition,This thesis also imposes a incoherence term on sub-dictionaries to ensure their independence,which further improves the discriminative ability of the learned dictionary.Similarly,given a query sample,the learned dictionary is first used for sparse coding,and then both the learned coding coefficients of training samples and the class-specific reconstruction errors are used for classification.The experimental results on the Extended Yale B,CMU PIE,AR and our real-world crop leaf disease datasets demonstrate the effectiveness and robustness of the proposed two dictionary learning methods.