Scholarlink

<p>Cross-modal image-text retrieval has gained increasing importance with the proliferation of multi-modal content on the Internet, which poses significant challenges in retrieving relevant images from text queries and vice versa. In recent years, knowledge distillation methods, leveraging large-scale pre-trained models to guide cross-modal retrieval models, have shown promising results. However, traditional distillation approaches often neglect the issue of knowledge forgetting during training, where the knowledge transferred from the pre-trained model to the student model diminishes over time, resulting in suboptimal performance. To address this limitation, we propose a dual-stage hash-aware distillation (DSHD) strategy that enhances retrieval performance by integrating pre-trained model distillation with hash-aware distillation and bilateral contrastive distillation (BCD), aiming to align features and hash codes better. Additionally, we employ a custom activation function and quantization strategy to reduce quantization error and minimize information loss during the feature-to-hash transformation. Extensive experiments on cross-modal hashing datasets demonstrate significant improvements over existing unsupervised hashing methods, validating the effectiveness of our approach. Our code is publicly available at: <a href="https://github.com/youonly-once/FDHD.">https://github.com/youonly-once/FDHD.</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

Dual-stage hash-aware distillation for unsupervised cross-modal hashing retrieval

  • 期刊论文

  • 发表日期：

    2025年8月6日

查看全文
  • Xinsheng Shu,
  • Ruiheng Li,
  • Mingyong Li

摘要

Cross-modal image-text retrieval has gained increasing importance with the proliferation of multi-modal content on the Internet, which poses significant challenges in retrieving relevant images from text queries and vice versa. In recent years, knowledge distillation methods, leveraging large-scale pre-trained models to guide cross-modal retrieval models, have shown promising results. However, traditional distillation approaches often neglect the issue of knowledge forgetting during training, where the knowledge transferred from the pre-trained model to the student model diminishes over time, resulting in suboptimal performance. To address this limitation, we propose a dual-stage hash-aware distillation (DSHD) strategy that enhances retrieval performance by integrating pre-trained model distillation with hash-aware distillation and bilateral contrastive distillation (BCD), aiming to align features and hash codes better. Additionally, we employ a custom activation function and quantization strategy to reduce quantization error and minimize information loss during the feature-to-hash transformation. Extensive experiments on cross-modal hashing datasets demonstrate significant improvements over existing unsupervised hashing methods, validating the effectiveness of our approach. Our code is publicly available at: https://github.com/youonly-once/FDHD.