Research On Image Classification Method Based On Multi-layer Dictionary Learning

With the leap development of modern information technology,more intelligent and digital lifestyle has gradually gained around and enriched people's life experience,which has become an inseparable normal life in the Internet era.How to effectively classify and utilize image data with different characteristics become a hot topic for researchers at home and abroad.However,the main characteristics of those data include the large volumes,low density of information and complex structural features,which is powerless to deal the images with conventional classification methods directly.According to the available research,the key to successful image classification is the extraction and screening of potentially discriminative features of the image data.Therefore,in order to settle this problem,this thesis,combine with the current research trend of multi-layer dictionary learning,try to study the discriminative feature extraction to realize the image classification efficiently and accurately.The main contribution of this paper is can be summarized as follow:Aiming at the problem that the redundancy of image features is hard to extract discriminant features,a multi-layer dictionary learning method based on label consistency is proposed,which is an extension of the traditional dictionary learning framework.The hierarchical dictionary learning and sparse coding are used to realize the inherited nonlinear projection transformation of the original image features,and the deep features in the image data are excavated in depth.Considering the independence of multi-layer dictionary learning framework model and classifier training,the discriminant label constraint items are integrated with multi-level learning,and the joint cooperative training optimization of training dictionary,sparse coding matrix and classifier parameters is realized.The feature subsets that are beneficial to improve the classification effect are screened out,and the differences among coding coefficient features of different categories are enhanced.Aiming at the problems of complex structural features and difficult extraction of local structural features from image data,a multi-layer dictionary learning method based on label embedding and locality constraint is proposed.Different from the existing methods,which only consider the discriminant information in the training samples,the proposed method achieves further mining of the discriminant features of atoms by constructing both multi-layer embedded constraints and multi-layer local structural constraints of atoms at the same time,and reduces the influence of noise in the original image.The discriminability of the feature coding matrix is enhanced while the atomic structure features remain unchanged.Finally,the discriminant feature transfer under the framework of multi-layer dictionary learning is realized by constructing the minimum error term of atomic label embedding constraint and atomic local feature constraint,so as to improve the classification performance of the overall model.Finally,experimental verification is carried out on several public image databases(such as Caltech101,LFW,etc.),and the experimental results are analyzed from multiple dimensions such as convergence verification,parameter sensitivity,training complexity,etc.,with the image classification accuracy as the measurement indicator.By comparing with the most representative image classification method based on the single-layer dictionary learning model and the deep learning method,it is shown that the image data classification method described in this paper has a remarkable performance in both the overall performance of classification and the training complexity,and can effectively adapt to the classification problems of various image data.