Scholarlink

<p>With the rapid development of multimedia technologies and social networking platforms, massive volumes of multi-modal data are being generated, increasing the demand for efficient information retrieval. To improve retrieval efficiency and better serve users in daily life and work, cross-modal retrieval has become a key research focus. Although deep cross-modal hashing methods have made significant progress, several challenges remain, including: 1) Insufficient expressive power of single-granularity semantic learning. 2) Lack of precise cross-modal alignment. 3) Missing or redundant use of explicit semantic labels. This paper proposes a novel method called Multi-Granularity Alignment Proxy Hashing (MAPH) to address these challenges. MAPH is designed to mine semantic concepts at multiple granularities while aligning heterogeneous modalities at the conceptual level for multi-modal, multi-granularity hash learning. Specifically, MAPH constructs a unified learning framework based on global-region-local multi-granularity features, enabling fine-grained semantic modeling via hierarchical feature extraction. Additionally, a contrastive learning-based multi-modal alignment module bridges the semantic gap between modalities and improves semantic consistency. MAPH also introduces a pairwise learning mechanism to optimize distance relationships between samples and incorporates a category proxy loss to enhance label semantic utilization, thus generating high-quality hash codes. Experiments on benchmark datasets show that MAPH significantly outperforms advanced methods in retrieval performance, validating its effectiveness.</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-granularity alignment proxy hashing for cross-modal retrieval

  • 期刊论文

  • 发表日期：

    2026年4月18日

查看全文
  • Jiayi Yang,
  • Xiuyuan Li,
  • Qinze Zhu,
  • Mingyong Li

摘要

With the rapid development of multimedia technologies and social networking platforms, massive volumes of multi-modal data are being generated, increasing the demand for efficient information retrieval. To improve retrieval efficiency and better serve users in daily life and work, cross-modal retrieval has become a key research focus. Although deep cross-modal hashing methods have made significant progress, several challenges remain, including: 1) Insufficient expressive power of single-granularity semantic learning. 2) Lack of precise cross-modal alignment. 3) Missing or redundant use of explicit semantic labels. This paper proposes a novel method called Multi-Granularity Alignment Proxy Hashing (MAPH) to address these challenges. MAPH is designed to mine semantic concepts at multiple granularities while aligning heterogeneous modalities at the conceptual level for multi-modal, multi-granularity hash learning. Specifically, MAPH constructs a unified learning framework based on global-region-local multi-granularity features, enabling fine-grained semantic modeling via hierarchical feature extraction. Additionally, a contrastive learning-based multi-modal alignment module bridges the semantic gap between modalities and improves semantic consistency. MAPH also introduces a pairwise learning mechanism to optimize distance relationships between samples and incorporates a category proxy loss to enhance label semantic utilization, thus generating high-quality hash codes. Experiments on benchmark datasets show that MAPH significantly outperforms advanced methods in retrieval performance, validating its effectiveness.