Scholarlink

<p>In the field of image fusion, integrating thermal radiation from infrared images with structural details from visible images can enhance scene information and support high-level tasks such as target recognition. However, existing methods are difficult to handle degraded images under complex conditions. Therefore, the paper proposes an infrared and visible image fusion algorithm based on automatic semantic guidance. The algorithm combines Contrastive Language-Image Pre-training (CLIP) model and Mixture of Exports (MOE) network, and introduces low-rank (LoRA) fine-tuning strategy, which can adaptively handle the degradation problem in multimodal images and effectively improve the quality of image fusion. In our method, a multi-task head supervises the process, while a high-level vision task enhances fusion quality. Extensive experiments on public datasets demonstrate the effectiveness of our method in restoring fine details, maintaining semantic consistency, and enhancing visual perception. Our approach outperforms mainstream state-of-the-art methods, achieving superior results in key metrics such as AG, EN, and VIF. The implementation is available at: <a href="https://github.com/Wohaizainuli/SGAFusion">https://github.com/Wohaizainuli/SGAFusion</a> SGAFusion (GitHub).</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

SGAFusion: a semantic-guided adaptive fusion framework for infrared and degraded visible images

  • 期刊论文

  • 发表日期：

    2026年4月17日

查看全文
  • Peng Hu,
  • Junjie Ma,
  • Yi Li

摘要

In the field of image fusion, integrating thermal radiation from infrared images with structural details from visible images can enhance scene information and support high-level tasks such as target recognition. However, existing methods are difficult to handle degraded images under complex conditions. Therefore, the paper proposes an infrared and visible image fusion algorithm based on automatic semantic guidance. The algorithm combines Contrastive Language-Image Pre-training (CLIP) model and Mixture of Exports (MOE) network, and introduces low-rank (LoRA) fine-tuning strategy, which can adaptively handle the degradation problem in multimodal images and effectively improve the quality of image fusion. In our method, a multi-task head supervises the process, while a high-level vision task enhances fusion quality. Extensive experiments on public datasets demonstrate the effectiveness of our method in restoring fine details, maintaining semantic consistency, and enhancing visual perception. Our approach outperforms mainstream state-of-the-art methods, achieving superior results in key metrics such as AG, EN, and VIF. The implementation is available at: https://github.com/Wohaizainuli/SGAFusion SGAFusion (GitHub).