摘要
目前的跨模态哈希检索方法通常假设不同模态数据与对应的语义标签内容完全匹配,但不同模态数据内在的特性与人工标注可能引入的噪声之间往往无法实现完全匹配。针对这个问题,提出了一种基于标签驱动语义感知学习的跨模态哈希检索(label-driven semantic-aware learning for cross-modal Hashing retrieval,LSLCHR)方法。通过图卷积网络(graph convolutional network,GCN)将标签信息补充至样本中,用于弥补样本中缺失的语义信息;将多标签作为引导信息,过滤掉样本中的非语义信息;通过一种量化差异性保持损失,减少了哈希码二值化过程中导致的样本相似性变化。在三个基线数据集上的大量实验表明,LSLCHR可以取得令人满意的检索效果。
Hashing techniques have been widely applied in cross-modal retrieval due to their advantages in fast and efficient retrieval.Existing cross-modal hashing retrieval methods typically assume that data from different modalities perfectly align with their corresponding semantic labels.However,this assumption often fails due to the inherent characteristics of different modalities and noise introduced by manual labeling,resulting in imperfect alignment between them.To address this issue,this paper proposed a label-driven semantic-aware learning for cross-modal Hashing retrieval(LSLCHR).First,a graph convolutional network(GCN)is utilized to integrate label information into the samples,compensating for missing semantic information.Second,multi-label is used as guiding information to filter out non-semantic information from the samples.Furthermore,a quantization discrepancy preservation loss is introduced to reduce the change in sample similarity caused by the binarization process of hash codes.Extensive experiments on three benchmark datasets demonstrate that the proposed LSLCHR method achieves satisfactory retrieval performance.
作者
师广田
张峰
张辉
朱杰
SHI Guangtian;ZHANG Feng;ZHANG Hui;ZHU Jie(College of Mathematics and Information Science,Hebei University,Baoding 071002,P.R.China)
出处
《重庆邮电大学学报(自然科学版)》
北大核心
2025年第6期870-883,共14页
Journal of Chongqing University of Posts and Telecommunications(Natural Science Edition)
基金
国家自然科学基金项目(61802269)
河北省自然科学基金项目(F2022511001)
河北省高等学校科学技术研究项目(ZC2022070)
河北大学高层次人才科研启动项目(521100223212)。
关键词
跨模态检索
深度哈希
图卷积网络
信息补充
信息过滤
cross-modal retrieval
deep Hashing
graph convolutional network
information supplementation
information filtering
