Scholarlink

<p>Reliable perception systems are crucial for the stable operation of applications such as autonomous driving and night-time security under extreme lighting conditions, including darkness and low light. The environmental perception modules within these systems typically rely on the precise segmentation of single-frame images. However, Existing RGB-infrared segmentation models often overlook modality discrepancies in feature fusion and suffer from detail loss due to repeated downsampling. To address these issues, we propose CMPFNet (Cross-modal Phased Fusion Net), a semantic segmentation model designed for extreme lighting conditions. CMPFNet employs a two-branch structure to capture local and global features, enhancing the segmentation accuracy of small targets. We introduce a feature rectification and fusion module to mitigate inter-modality gaps and improve feature interaction via multimodal rectification and attention mechanisms. A feature recovery module is introduced to restore fine details lost in deep layers, while a multi-scale decoding module reduce noise interference across scales. Experiments on MFNet and PST900 datasets demonstrate that CMPFNet outperforms mainstream methods, achieving a 60.6% mIoU on MFNet, a 1.7% improvement over the previous best method, validating its effectiveness. This approach provides a technical foundation for robust environmental perception in the above-mentioned scenarios. Code is available at <a href="https://github.com/tzh0831/CMPFNet">https://github.com/tzh0831/CMPFNet</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

CMPFNet: semantic segmentation network for cross-modal phased fusion in extreme light scenes

  • 期刊论文

  • 发表日期：

    2025年10月28日

查看全文
  • Jing Han,
  • Zihe Tian,
  • Koukou Gao,
  • Xueqiang Lyu

摘要

Reliable perception systems are crucial for the stable operation of applications such as autonomous driving and night-time security under extreme lighting conditions, including darkness and low light. The environmental perception modules within these systems typically rely on the precise segmentation of single-frame images. However, Existing RGB-infrared segmentation models often overlook modality discrepancies in feature fusion and suffer from detail loss due to repeated downsampling. To address these issues, we propose CMPFNet (Cross-modal Phased Fusion Net), a semantic segmentation model designed for extreme lighting conditions. CMPFNet employs a two-branch structure to capture local and global features, enhancing the segmentation accuracy of small targets. We introduce a feature rectification and fusion module to mitigate inter-modality gaps and improve feature interaction via multimodal rectification and attention mechanisms. A feature recovery module is introduced to restore fine details lost in deep layers, while a multi-scale decoding module reduce noise interference across scales. Experiments on MFNet and PST900 datasets demonstrate that CMPFNet outperforms mainstream methods, achieving a 60.6% mIoU on MFNet, a 1.7% improvement over the previous best method, validating its effectiveness. This approach provides a technical foundation for robust environmental perception in the above-mentioned scenarios. Code is available at https://github.com/tzh0831/CMPFNet.