Scholarlink

<p>Infrared and visible image fusion is a pivotal task in computer vision, aiming to integrate complementary modal characteristics to generate fused images suitable for both human visual observation and machine perception. However, most state-of-the-art fusion algorithms prioritize visual quality at the expense of utility for downstream high-level vision tasks, and existing semantic-driven methods face critical limitations: noise aliasing from coarse frequency-domain processing and gradient conflicts in multi-task joint training of fusion and segmentation. To address these challenges, this paper proposes ViR-MFS (Visible and Infrared Image Meta-learning Framework for Fusion and Segmentation), an end-to-end joint network for infrared-visible image fusion and segmentation based on alternating meta-learning optimization and a multi-scale wavelet fusion module. First, a MixVisionTransformer is adopted as a shared backbone to construct a multi-scale feature pyramid that balances local spatial continuity and global semantic dependencies. Second, a Multi-scale Wavelet Fusion Module (MWFM) is designed to explicitly decouple high- and low-frequency components in the frequency domain; with learnable scale factors for adaptive recalibration, MWFM enables precise injection of visible texture details while effectively suppressing infrared thermal noise. To resolve objective conflicts in multi-task learning, an alternating meta-learning optimization strategy is introduced, which dynamically coordinates the training of fusion and segmentation tasks to guide the backbone in learning robust features with both fine texture details and strong semantic discriminability, mitigating inter-task performance trade-offs. Extensive experiments on the MSRS and FMB datasets demonstrate that ViR-MFS achieves superior performance in both fusion quality and downstream semantic segmentation. To enhance the transparency and reproducibility of our research, the source code and datasets, along with comprehensive usage guidelines, are publicly available at <a href="https://github.com/wyh13617290776/ViR-MFS">https://github.com/wyh13617290776/ViR-MFS</a> (archived with DOI: 10.5281/zenodo.18740397).</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

Wavelet-driven meta-learning: unifying infrared-visible fusion and semantic segmentation for robust scene perception

  • 期刊论文

  • 发表日期：

    2026年7月24日

查看全文
  • Shichao Liu,
  • Chun Sun,
  • Dengshi Li,
  • Yihui Wang,
  • Shiwei Hu,
  • Zhiming Zhan

摘要

Infrared and visible image fusion is a pivotal task in computer vision, aiming to integrate complementary modal characteristics to generate fused images suitable for both human visual observation and machine perception. However, most state-of-the-art fusion algorithms prioritize visual quality at the expense of utility for downstream high-level vision tasks, and existing semantic-driven methods face critical limitations: noise aliasing from coarse frequency-domain processing and gradient conflicts in multi-task joint training of fusion and segmentation. To address these challenges, this paper proposes ViR-MFS (Visible and Infrared Image Meta-learning Framework for Fusion and Segmentation), an end-to-end joint network for infrared-visible image fusion and segmentation based on alternating meta-learning optimization and a multi-scale wavelet fusion module. First, a MixVisionTransformer is adopted as a shared backbone to construct a multi-scale feature pyramid that balances local spatial continuity and global semantic dependencies. Second, a Multi-scale Wavelet Fusion Module (MWFM) is designed to explicitly decouple high- and low-frequency components in the frequency domain; with learnable scale factors for adaptive recalibration, MWFM enables precise injection of visible texture details while effectively suppressing infrared thermal noise. To resolve objective conflicts in multi-task learning, an alternating meta-learning optimization strategy is introduced, which dynamically coordinates the training of fusion and segmentation tasks to guide the backbone in learning robust features with both fine texture details and strong semantic discriminability, mitigating inter-task performance trade-offs. Extensive experiments on the MSRS and FMB datasets demonstrate that ViR-MFS achieves superior performance in both fusion quality and downstream semantic segmentation. To enhance the transparency and reproducibility of our research, the source code and datasets, along with comprehensive usage guidelines, are publicly available at https://github.com/wyh13617290776/ViR-MFS (archived with DOI: 10.5281/zenodo.18740397).