Scholarlink

<p>Current approaches to image–text retrieval face challenges in achieving fine-grained semantic alignment and scaling up negative samples due to the complexities of fine-grained correspondence learning and the reliance on mini-batches. To address these issues, we propose a novel method called Enhancing Cross-Modal Retrieval through Element-Level Semantic Enrichment and Momentum Contrast. First, we design a plug-and-play element-level semantic enhancement block that removes redundant regional features and assigns adaptive weights to enhance semantic expressiveness. Second, we incorporate a momentum memory module to dynamically store negative samples, effectively expanding the scale of negative samples at the batch level. Our mixed training objective function enables more comprehensive perception of the semantic information for each sample. Experiments on the MS-COCO and Flickr30K datasets demonstrate the effectiveness of our approach, which outperforms state-of-the-art methods. The code can be obtained from <a href="https://github.com/wangmeng-007/ECMR">https://github.com/wangmeng-007/ECMR</a>.</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

Enhancing cross-modal retrieval through element-level semantic enrichment and momentum contrast

  • 期刊论文

  • 发表日期：

    2025年12月10日

查看全文
  • Jiayi Fu,
  • Meng Wang,
  • Kaiqi Yang,
  • Hao Chen,
  • Guangyun Lu

摘要

Current approaches to image–text retrieval face challenges in achieving fine-grained semantic alignment and scaling up negative samples due to the complexities of fine-grained correspondence learning and the reliance on mini-batches. To address these issues, we propose a novel method called Enhancing Cross-Modal Retrieval through Element-Level Semantic Enrichment and Momentum Contrast. First, we design a plug-and-play element-level semantic enhancement block that removes redundant regional features and assigns adaptive weights to enhance semantic expressiveness. Second, we incorporate a momentum memory module to dynamically store negative samples, effectively expanding the scale of negative samples at the batch level. Our mixed training objective function enables more comprehensive perception of the semantic information for each sample. Experiments on the MS-COCO and Flickr30K datasets demonstrate the effectiveness of our approach, which outperforms state-of-the-art methods. The code can be obtained from https://github.com/wangmeng-007/ECMR.