Scholarlink

<p>Cross-modal hashing retrieval tasks have attracted much attention due to their fast retrieval speed and low storage cost. However, existing unsupervised methods still have two limitations. (1) Current methods fail to fully exploit the multidimensional information among samples, resulting in a suboptimal semantic space. (2) Mainstream methods often rest on the strong assumption of full cross-modal pairing. Yet in real applications, data collection uncertainty leads to sub-quality samples with semantic matching deviations. Because of the low storage density of hash codes, the introduction of such samples increases the information burden in the learning process of hash functions and reduces the precision of hash codes. To address the above problems, we propose a method called Hierarchical Multidimensional Sample Dynamic Optimization for Cross-Modal Hashing (UMOH). This method uses a multi-level feature decomposition and fusion mechanism. It lets the hash function fully capture the multi-level semantic correlations of cross-modal data. Furthermore, we dynamically adjust the training weights of samples to mitigate their negative impact on the hash function. Meanwhile, we leverage the loss value of high-level features to guide the optimization of hash loss. Extensive experiments conducted on multiple datasets demonstrate the superiority of our method in retrieval accuracy. The source code of our method can be accessed at <a href="https://github.com/dh2425/UMOH">https://github.com/dh2425/UMOH</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

Unsupervised multidimensional sample dynamic optimization for cross-modal hashing

  • 期刊论文

  • 发表日期：

    2026年4月15日

查看全文
  • Hao Du,
  • Li Liu,
  • Huaxiang Zhang,
  • Dongmei Liu,
  • Kang Du

摘要

Cross-modal hashing retrieval tasks have attracted much attention due to their fast retrieval speed and low storage cost. However, existing unsupervised methods still have two limitations. (1) Current methods fail to fully exploit the multidimensional information among samples, resulting in a suboptimal semantic space. (2) Mainstream methods often rest on the strong assumption of full cross-modal pairing. Yet in real applications, data collection uncertainty leads to sub-quality samples with semantic matching deviations. Because of the low storage density of hash codes, the introduction of such samples increases the information burden in the learning process of hash functions and reduces the precision of hash codes. To address the above problems, we propose a method called Hierarchical Multidimensional Sample Dynamic Optimization for Cross-Modal Hashing (UMOH). This method uses a multi-level feature decomposition and fusion mechanism. It lets the hash function fully capture the multi-level semantic correlations of cross-modal data. Furthermore, we dynamically adjust the training weights of samples to mitigate their negative impact on the hash function. Meanwhile, we leverage the loss value of high-level features to guide the optimization of hash loss. Extensive experiments conducted on multiple datasets demonstrate the superiority of our method in retrieval accuracy. The source code of our method can be accessed at https://github.com/dh2425/UMOH.