Scholarlink

<p>The registration of visible and infrared images is a critical task in multimodal image processing. However, geometric deformations and cross-modal feature discrepancies pose significant challenges, causing traditional feature-based methods like SIFT and SuperGlue to produce registration errors, particularly under extreme illumination changes or partial occlusions. To address these issues, we propose an edge-guided registration framework based on LoFTR alignment (EdgeLF), for infrared–visible alignment, which combines edge-guided coarse alignment with local binary pattern (LBP)-based texture refinement. Experimental results show that EdgeLF achieves subpixel registration precision, delivering up to 57.1% higher accuracy than LoFTR under large-scale deformations and severe illumination variations on the RoadScene and LLVIP datasets.</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

EdgeLF: edge-guided registration with loftr for visible and infrared images

  • 期刊论文

  • 发表日期：

    2025年12月11日

查看全文
  • Haicheng Zhu,
  • Cheng Zeng,
  • Yijin Pan,
  • Anzheng Tang,
  • Junbo Wang

摘要

The registration of visible and infrared images is a critical task in multimodal image processing. However, geometric deformations and cross-modal feature discrepancies pose significant challenges, causing traditional feature-based methods like SIFT and SuperGlue to produce registration errors, particularly under extreme illumination changes or partial occlusions. To address these issues, we propose an edge-guided registration framework based on LoFTR alignment (EdgeLF), for infrared–visible alignment, which combines edge-guided coarse alignment with local binary pattern (LBP)-based texture refinement. Experimental results show that EdgeLF achieves subpixel registration precision, delivering up to 57.1% higher accuracy than LoFTR under large-scale deformations and severe illumination variations on the RoadScene and LLVIP datasets.