Research On Discriminative Dictionary Learning Algorithm For Image Recognition

Dictionary learning is one of the important research contents in machine learning and related research fields.It has been widely used in image reconstruction,image denoising,image segmentation and other fields.And it has attracted the attention of researchers.The traditional dictionary learning focuses on the sparse representation learning of the sub-dictionary representation ability and the collaborative representation learning of all class dictionary representations.However,there are relatively few dictionary learning and multi-view dictionary learning studys under multiple constraints such as dimensionality reduction,and low rank and inconsistency.Therefore,this thesis focuses on dictionary learning and multi-view dictionary learning based on multiple constraints.The main research work is as follows:1.Propose a fast low-rank shared dictionary learning with sparsity constraint on face recognition(FLRSDLSC)algorithm.The algorithm adopts the strategy of simultaneous dimensionality reduction and dictionary learning.The main idea is as follows:the orthogonality of the projection matrix could be protected by Cayley transform to obtain compact features;the fisher discriminant criterion is applied to the class-specific dictionaries to improve the discriminative ability of each-class dictionary;the low-rank constraint is embedded in the shared class dictionary to reduce redundancy,which effectively enhances the discriminative ability of dictionary and coding coefficients.Experiments on four datasets(AR,Extended Yale B,CMU PIE and FERET datasets)validate the effectiveness of the FLRSDLSC algorithm.2.Introduce a multi-constraints-Based Enhanced Class-Specific Dictionary Learning(MECDL)algorithm.The algorithm divides the dictionary into class-specific dictionary and a shared dictionary,and these dictionaries are exerted low-rank constraints.For the class-specific dictionaries,we simultaneously apply fisher's discriminant criterion and low-rank constraint into it to obtain a compact class-specific sub-dictionary,which enhances the discriminative ability of the class-specific sub-dictionary.For the shared dictionary,we embed low-rank constraint to reduce the redundancy of it.Experiments on three datasets(AR,Extended Yale B,and COIL-20 datasets)validate the effectiveness of the MECDL algorithm.3.Present two multi-view dictionary learning algorithms based on collaborative representation:Multi-View Dictionary Learning based on intra-view atom inconsistency algorithm(MDLIAI)and Multi-View Dictionary Learning based on inter-view dictionary inconsistency algorithm(MDLIDI).The common characteristics of two algorithms are:they learn the class-specific dictionary and a shared class dictionary for each view,and introduce the minimum of the coding coefficients variance to reduce the diversity of inter-view dictionaries;in addition,the minimization of the weighted sum of the distance between the each-view coding coefficients and the mean of all-view coding coefficients restricts the contribution of the corresponding features.The difference of two algorithms is:MDLIAI embeds the inconsistency constraint into the intra-view dictionary,but the method of updating dictionary spends lots of time at the atomic level.To improve this problem,MDLIDI embeds the inconsistency constraints into the inter-view dictionary.Experiments on two datasets(AR and Extended Yale B datasets)verify the effectiveness of the MDLIDI and MDLIAI algorithms.