Font Size: a A A

Research On Supervised Multi-view Hashing Algorithms For Image Retrieval

Posted on:2022-11-26Degree:MasterType:Thesis
Country:ChinaCandidate:R J GaoFull Text:PDF
GTID:2518306611494714Subject:Computer Software and Application of Computer
Abstract/Summary:PDF Full Text Request
With the rapid development of the Internet and data storage technology,multimedia data is constantly emerging,among which image data can provide intuitive and rich visual information,so it has become a research hotspot.In image retrieval,image feature extraction,similarity matching and large-scale indexing are always the research focus in the fields of machine learning,computer vision and information retrieval.Hashing technology has received extensive attention in multimedia retrieval due to the advantages of fast retrieval speed and low storage cost.The multi-view hashing algorithms for image retrieval aim to extract features from image data and map them to discrete Hamming spaces for fast approximate nearest neighbor retrieval.The existing multi-view hashing algorithms consider the fusion of multiple view features,and map multi view features into discrete binary codes.However,some multi-view hashing algorithms still have shortcomings in maintaining the similarity of data.Some multi-view hashing algorithms cause large quantization errors in the discrete optimization stage of data,or lead to the problem of high time complexity.This paper focuses on the supervised multi-view hashing algorithms for image retrieval task,with the main goal of integrating multi-view features and learning compact binary hashing codes.The main research contents and innovations of this paper are as follows:(1)Fractional Multi-view Hashing with Semantic Correlation Maximization(FMH-SCM)algorithm is proposed.The existing unsupervised multi-view hashing algorithms do not use supervised information and the learned hashing codes do not have strong discrimination.FMH-SCM not only considers learning the multi-view hashing codes with labels to maximize the semantic correlation between different views.Moreover,the noise interference in the data is also considered and the fractional order embedding idea is adopted to reduce the negative influence of noise on the data through fractional order modeling.Experimental results verify the effectiveness of FMH-SCM algorithm.(2)Discrete Multi-view Hashing with Fine Semantics(DMVH-FS)algorithm is proposed.This algorithm reconsiders the construction of similarity matrix.After non-linear embedding of the data,the fine semantics between samples is learned by maintaining the semantic similarity within and between classes,so as to construct a similarity matrix with fine semantics.In the optimization solution,a new discrete hash optimization method is used to reduce the quantization error without relaxing the discrete constraints.Experiments on real datasets show that DMVH-FS algorithm is superior to existing hashing algorithms.(3)Self-Adaptive Weighted Unmediated Multi-view Hashing(SAWUMH)algorithm is proposed.The algorithm learns multi-view hashing codes directly from the labels without adopting any intermediate representation,which can also reduce the number of variables required in the optimization process.At the same time,in order to improve the discriminativeness of hashing codes,SAWUMH uses semantic labels and fine semantic similarity matrix as double supervision in the process of hashing learning to improve the accuracy.In addition,according to the contribution of each view,SAWUMH adopts an adaptive weighting method to assign reasonable weights to each view.Experimental results show that SAWUMH algorithm has better retrieval performance.
Keywords/Search Tags:Image retrieval, Hashing learning, Discrete binary optimization, Multiview learning
PDF Full Text Request
Related items