Scholarlink

<p>Zero-shot learning (ZSL) enables recognition of unseen classes without direct training examples by leveraging semantic knowledge. However, traditional ZSL methods often rely on static, expert-defined semantics, leading to misalignment with visual features and degraded performance, especially on fine-grained datasets. To address this, we propose RSDN-GZSL, a recurrent semantic disentanglement network that dynamically generates instance-specific semantic information for precise semantic–visual alignment and effective feature decoupling. Our model comprises a dual-path feature generation network and an evolution attention network, forming an iterative optimization loop. Extensive experiments on AWA2, SUN, and CUB datasets demonstrate that RSDN-GZSL significantly outperforms state-of-the-art methods, establishing new performance records. The code is available at <a href="https://github.com/JingHu-gdut/RSDN-GZSL">https://github.com/JingHu-gdut/RSDN-GZSL</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

Recurrent semantic disentanglement: enhancing zero-shot learning through dynamic feature refinement

  • 期刊论文

  • 发表日期：

    2026年3月22日

查看全文
  • Jing Hu,
  • Min Meng,
  • Jigang Wu

摘要

Zero-shot learning (ZSL) enables recognition of unseen classes without direct training examples by leveraging semantic knowledge. However, traditional ZSL methods often rely on static, expert-defined semantics, leading to misalignment with visual features and degraded performance, especially on fine-grained datasets. To address this, we propose RSDN-GZSL, a recurrent semantic disentanglement network that dynamically generates instance-specific semantic information for precise semantic–visual alignment and effective feature decoupling. Our model comprises a dual-path feature generation network and an evolution attention network, forming an iterative optimization loop. Extensive experiments on AWA2, SUN, and CUB datasets demonstrate that RSDN-GZSL significantly outperforms state-of-the-art methods, establishing new performance records. The code is available at https://github.com/JingHu-gdut/RSDN-GZSL.