Research On Image Attribute Multi-label Classification And Retrieval

As an international frontier research in the field of information,big data processing is becoming an important support for the intelligent development of the Internet,cloud computing,and the Internet of Things.As the key technologies of big data processing,attribute multilabel classification and image retrieval greatly improve the quality and efficiency of image data classification and query.However,low discriminant problems such as semantic abstraction of attribute and same attribute of foreign objects lead to poor classification results,and poor discriminant problems such as image scale change and perspective difference lead to low image retrieval performance.How to accurately identify attributes and represent images,and achieve accurate attribute multi-label classification and image retrieval has become a key problem to be solved urgently at present.This thesis thorough analyzes and summarizes present development situation and the existing difficulties and challenges of the image attributes classification and image retrieval technology.Based on deep learning algorithm,a space and semantic information associated attributes discriminant model is constructed and a method of attribute feature location and reconstruction is proposed,which can achieve accurate classification of attributes.Besides,this thesis builds an image representation framework for images of class differentiation and proposes an accurate sorting method of image similarity to improve the performance of image retrieval algorithm.The specific research contents are as follows:(1)In view of the problem that the multi-label classification methods of instance-oriented attributes relies too much on labelled regions,the second chapter analyzes the region mapping principle and constructs a region prediction learnable framework of attributes.And the multilabel classification and location method of instance-oriented attributes based on global-guided weakly supervised learning is further proposed.The method disentangles features through the estimated global correlation and explores the related fine-grained features to obtain accurate label area prediction,which can construct of the mapping relationship between image-level labels and related areas and improve precision of label localization and classification.Compared with the current advanced correlation methods,label localization is more accurate and label recognition performance is better.(2)Similar properties of different objects of categories may have great differences in visual appearance for concept-oriented attributes,the third chapter analyzes characteristics and classification mechanism of the concept-oriented attributes and proposes self-supervised deconstruction and reconstruction learning method for the concept-oriented attributes classification.In this method,more fine-grained local features are obtained through self-supervised image deconstruction and feature reconstruction,and end-to-end multi-task complementary optimization greatly improves the multi-label classification accuracy of concept-oriented attributes and solves the problem of ”same attribute of foreign objects”.The experimental performance is obviously improved compared with the current research level.(3)Aiming at solving the problems of scale variation and low correlation in the process of single label image retrieval,the consistency characterization framework and feature correlation calculation model of different scale features are constructed in chapter four.An image retrieval method based on end-to-end skip-connected network and Gram matrix is proposed.In this method,multi-resolution features of different layers are extracted by skip-connected manner to solve the problem of scale change,and the second-order information of features is calculated by Gram matrix to explore their correlation.Based on the joint learning of similar information and semantic information between images,hash coding is end-to-end optimized to achieve in-class clustering and inter-class dispersion in Hamming space and improve the coding quality.Compared with the existing retrieval technology,the retrieval accuracy is improved obviously.(4)Traditional binary similarity definition cannot describe similarity sorting of the different number of shared labels,the fifth chapter analyzes the corresponding relationship between the number of shared labels and similarity definition and puts forward label-guided similarity loss function,which achieves accurate sorting of the similarity of relatively similar images.At the same time,the feature decomposition and reciprocal learning model are constructed in this chapter to further improve the ability of feature representation and enhance the ability to distinguish retrieved samples,so as to specifically solve the problem of accurate similarity sorting of samples of different correlations for multi-label data.Compared with the current research level,the retrieval accuracy and sorting ability are better.