Scholarlink

<p>Visible and infrared image fusion (VIF) is essential for unmanned aerial vehicle (UAV) perception by integrating complementary information from heterogeneous sensors to enhance understanding in complex environments. Existing deep learning-based VIF methods often face a trade-off between the limited receptive fields of CNNs and the high computational overhead of Transformers. Although Mamba shows great potential in modeling long-range dependencies with linear complexity, it struggles to preserve spatial information in complex spatial correlations. Moreover, most VIF methods lack effective multimodal feature disentanglement, which is particularly challenging in low-light and cluttered backgrounds. To address these challenges, we propose MCDFuse, a novel Mamba-CNN-driven VIF framework that synergistically combines the strengths of Mamba and CNN architectures. We introduce a dynamic feature activation module to improve spatial modeling and facilitate high-quality image reconstruction. A dual-branch architecture integrates ConvNeXt modules to expand the receptive field and inverted residual blocks to enhance detail representation. Furthermore, a feature decomposition loss with normalization constraints is designed to optimize feature disentanglement during training. Comparative experiments on multiple datasets, including newly constructed ones, against 12 state-of-the-art methods demonstrate that our approach achieves superior fusion quality and improves performance on downstream object detection tasks. The dataset can be downloaded from <a href="https://github.com/MCDFuse/UAVScene">https://github.com/MCDFuse/UAVScene</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

MCDFus: visible and infrared image fusion via Mamba-CNN for enhanced UAV perception

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2025年12月4日

查看全文
  • Yu Zhang,
  • Hao Cha,
  • HanRui Zhang,
  • XiongHui Li,
  • Zhuo Chen,
  • JiaMing Bai,
  • TingTing Fu,
  • Yong Cheng Du,
  • Xian Zhang

摘要

Visible and infrared image fusion (VIF) is essential for unmanned aerial vehicle (UAV) perception by integrating complementary information from heterogeneous sensors to enhance understanding in complex environments. Existing deep learning-based VIF methods often face a trade-off between the limited receptive fields of CNNs and the high computational overhead of Transformers. Although Mamba shows great potential in modeling long-range dependencies with linear complexity, it struggles to preserve spatial information in complex spatial correlations. Moreover, most VIF methods lack effective multimodal feature disentanglement, which is particularly challenging in low-light and cluttered backgrounds. To address these challenges, we propose MCDFuse, a novel Mamba-CNN-driven VIF framework that synergistically combines the strengths of Mamba and CNN architectures. We introduce a dynamic feature activation module to improve spatial modeling and facilitate high-quality image reconstruction. A dual-branch architecture integrates ConvNeXt modules to expand the receptive field and inverted residual blocks to enhance detail representation. Furthermore, a feature decomposition loss with normalization constraints is designed to optimize feature disentanglement during training. Comparative experiments on multiple datasets, including newly constructed ones, against 12 state-of-the-art methods demonstrate that our approach achieves superior fusion quality and improves performance on downstream object detection tasks. The dataset can be downloaded from https://github.com/MCDFuse/UAVScene.