Scholarlink

<p>Multimodal image fusion aims to integrate complementary information from various imaging modalities to produce high-quality fused images with extensive global semantics and intricate detail textures. Traditional methods often rely on Transformers and CNNs, which face computational bottlenecks with high-resolution images. This paper introduces FocalNetFuse, a novel approach employing Focal Modulation Networks (FMNs) to efficiently capture long-range features and Invertible Neural Networks (INNs) to extract short-range details. Our mutual information-driven feature decomposition strategy enhances global consistency while preserving modality-specific details. The dual-level fusion architecture, comprising global and local fusion layers, demonstrates promising results in both infrared–visible light and medical image fusion tasks, achieving superior performance across multiple evaluation metrics. Here we demonstrate that our method achieves a notable improvement in entropy and standard deviation metrics compared to state-of-the-art methods, highlighting its potential for enhancing image quality and detail retention. This approach sets a new benchmark in the field of image fusion and opens up new avenues for the application of advanced neural network architectures in multimodal image processing. The code available at <a href="https://github.com/Bing998666/FocalNetFuse.git">https://github.com/Bing998666/FocalNetFuse.git</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

FocalNetFuse: enhancing multimodal image fusion quality with focal modulation networks

  • 期刊论文

  • 发表日期：

    2026年1月7日

查看全文
  • Bing Zou,
  • Xiaoliang Zhu,
  • Xusheng Wang,
  • Yangxun Liu,
  • Shuang Zhu,
  • Mengke Wen,
  • Ququn Shen

摘要

Multimodal image fusion aims to integrate complementary information from various imaging modalities to produce high-quality fused images with extensive global semantics and intricate detail textures. Traditional methods often rely on Transformers and CNNs, which face computational bottlenecks with high-resolution images. This paper introduces FocalNetFuse, a novel approach employing Focal Modulation Networks (FMNs) to efficiently capture long-range features and Invertible Neural Networks (INNs) to extract short-range details. Our mutual information-driven feature decomposition strategy enhances global consistency while preserving modality-specific details. The dual-level fusion architecture, comprising global and local fusion layers, demonstrates promising results in both infrared–visible light and medical image fusion tasks, achieving superior performance across multiple evaluation metrics. Here we demonstrate that our method achieves a notable improvement in entropy and standard deviation metrics compared to state-of-the-art methods, highlighting its potential for enhancing image quality and detail retention. This approach sets a new benchmark in the field of image fusion and opens up new avenues for the application of advanced neural network architectures in multimodal image processing. The code available at https://github.com/Bing998666/FocalNetFuse.git.