Scholarlink

<p>Infrared–visible image fusion (IVIF) aims to integrate complementary information from different modalities while preserving both structural consistency and fine-grained details. However, existing methods often struggle to simultaneously model modality-shared consistency and modality-specific complementarity. To address this issue, we propose ACDFuse, an Adaptive Cross-Modality Dual-Branch Fusion Network that decouples shared and specific representations for more effective cross-modal interaction. The proposed framework adopts a dual-branch encoder composed of a Serially Enhanced Representation Transformer Encoder for global consistency modeling and an invertible neural network encoder for preserving modality-specific details. Furthermore, adaptive fusion mechanisms, including an Adaptive Position-aware Cross-Attention module and a Hierarchical Fusion Layer, are introduced to achieve accurate cross-modal alignment and progressive feature integration. Extensive experiments on multiple public datasets demonstrate that ACDFuse achieves superior performance compared with existing state-of-the-art methods in both qualitative and quantitative evaluations, while also improving downstream object detection performance.The source code and trained models are available at <a href="https://github.com/730915/ACDFuse">https://github.com/730915/ACDFuse</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

Adaptive dual-branch fusion: enhancing infrared-visible image integration via cross-modal interaction

  • 期刊论文

  • 发表日期：

    2026年6月2日

查看全文
  • Huan Yu,
  • Xiaohong Wang

摘要

Infrared–visible image fusion (IVIF) aims to integrate complementary information from different modalities while preserving both structural consistency and fine-grained details. However, existing methods often struggle to simultaneously model modality-shared consistency and modality-specific complementarity. To address this issue, we propose ACDFuse, an Adaptive Cross-Modality Dual-Branch Fusion Network that decouples shared and specific representations for more effective cross-modal interaction. The proposed framework adopts a dual-branch encoder composed of a Serially Enhanced Representation Transformer Encoder for global consistency modeling and an invertible neural network encoder for preserving modality-specific details. Furthermore, adaptive fusion mechanisms, including an Adaptive Position-aware Cross-Attention module and a Hierarchical Fusion Layer, are introduced to achieve accurate cross-modal alignment and progressive feature integration. Extensive experiments on multiple public datasets demonstrate that ACDFuse achieves superior performance compared with existing state-of-the-art methods in both qualitative and quantitative evaluations, while also improving downstream object detection performance.The source code and trained models are available at https://github.com/730915/ACDFuse.