Scholarlink

<p>A significant semantic gap persists in cross-modal image–text matching due to disparities in abstraction levels, symbolic representations, and information modalities. Current scene graph-based approaches, which primarily rely on local feature alignment, face three key limitations: unimodal graph construction with implicit cross-modal correlations, insufficient global semantic modeling, and model drift caused by negative sample training. To address these challenges, this paper proposes a multi-scale feature and historical attention model (MSFHA). The framework extracts hierarchical global semantic representations from visual and textual data through multi-scale feature extraction. It incorporates an adaptively weighted hierarchical feature aggregation strategy into graph attention structures, and synergizes multiple attention mechanisms to achieve deep fusion of cross-modal graph features. Innovatively, a historical attention preservation mechanism is introduced to suppress overfitting during negative sample training. Evaluations on MSCOCO and Flickr30K benchmarks demonstrate comprehensive performance improvements of approximately 2% and 1%, respectively.\query{Please check the edit made in the article title.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Multi-scale feature and historical attention-based cross-modal image–text matching model

  • 期刊论文

  • 发表日期：

    2026年1月24日

查看全文
  • Liqin Wang,
  • Jiayi Liu,
  • Pengcheng Yang,
  • Yongfeng Dong,
  • Xu Wang

摘要

A significant semantic gap persists in cross-modal image–text matching due to disparities in abstraction levels, symbolic representations, and information modalities. Current scene graph-based approaches, which primarily rely on local feature alignment, face three key limitations: unimodal graph construction with implicit cross-modal correlations, insufficient global semantic modeling, and model drift caused by negative sample training. To address these challenges, this paper proposes a multi-scale feature and historical attention model (MSFHA). The framework extracts hierarchical global semantic representations from visual and textual data through multi-scale feature extraction. It incorporates an adaptively weighted hierarchical feature aggregation strategy into graph attention structures, and synergizes multiple attention mechanisms to achieve deep fusion of cross-modal graph features. Innovatively, a historical attention preservation mechanism is introduced to suppress overfitting during negative sample training. Evaluations on MSCOCO and Flickr30K benchmarks demonstrate comprehensive performance improvements of approximately 2% and 1%, respectively.\query{Please check the edit made in the article title.