Research On Label Hierarchy Based Cross-Modal Online Hashing

With the increasing trend of multimedia data pouring into the Internet,how to conduct efficient and accurate information retrieval has become a significant challenge that people must face.In the study of cross-modal retrieval methods,we must consider the semantic gap caused by the inconsistency of data characteristics and semantic information,and the heterogeneous gap faced by the retrieval between different modal data.Hash search is a typical nearest neighbor search technique that solves semantic and heterogeneous gap problems.By mapping from high-dimensional feature space to low-dimensional Hamming space,hash retrieval can effectively solve the problems of high data dimensions and high computation costs in retrieving large-scale multimedia data sets.However,the existing cross-modal online hashing methods ignore the label hierarchy.In addition,the hash function changes in the online scene cause the hash code of old data to become invalid,the representation and preservation of label hierarchy information,and the hash code optimization under the online hash learning framework pose new challenges to cross-modal online hashing.To solve the above problems,the work carried out in this thesis mainly includes the following:(1)This thesis has applied hierarchical labels to the online hash field to improve the performance of cross-modal retrieval for the first time,and has designed a unified online hash learning framework,which mainly includes hash code learning and hash function learning.(2)To address the issue of representing label hierarchy information,this thesis has proposed a soft hierarchical label based on the hierarchical structure for the first time,which can more accurately represent fine-grained semantic information between instances and categories.(3)To address the problem of invalid hash codes for old data and the preservation of label hierarchy information,this thesis has proposed the hierarchical virtual center based on category target codes.In contrast to randomly generated target codes,the hierarchical virtual center can preserve cross-layer association information between categories.(4)To address the hash code optimization problem in the online hash code learning framework,this paper has proposed an efficient discrete optimization algorithm that avoids the problem of semantic information loss and retrieval performance degradation caused by a relaxation strategy.This algorithm can efficiently learn hash codes and solve the hash code optimization problem in the online hash learning framework.(5)The thesis has conducted extensive experiments on two publicly available hierarchical datasets.The experimental results have shown that the proposed method outperforms existing methods in retrieval performance.On the FashionVC dataset,the MAP results of text retrieval by image query have been improved by an average of 2.24%,and an average of 1.10% has improved the MAP results of image retrieval by text query.