Scholarlink

<p>This study focuses on intelligent detection of groundwater seepage states on drill-and-blast tunnel faces and establishes a unified evaluation framework spanning two modalities—visible (VIS) and infrared (IR)—and five models (Faster R-CNN, SSD, DETR, YOLOv8, YOLOv11). VIS and IR images were sequentially acquired during the same inspection sessions on the same tunnel faces at four ongoing tunnel sites in western China, and, in accordance with the Railway Tunnel Design Code (TB 10003 − 2016) and on-site flow measurements, three seepage states—damp/dripping, rain-like/linear, and flowing—were annotated. All models were compared under identical data splits and unified training/inference protocols, with evaluation metrics including mAP@50, mAP@50–95, training-dynamics curves, and FPS. The comparative results show that, relative to VIS, IR provides higher overall stability and better recognition of safety-critical classes (“rain-like/linear” and “flowing”). For engineering deployment, the YOLO family offers a more favorable accuracy–speed trade-off and is therefore recommended as the primary choice.</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

Comparing Infrared and Visible Imaging for Tunnel Groundwater Recognition: A Multi-Model Systematic Evaluation

  • 期刊论文

  • 发表日期：

    2026年5月29日

查看全文
  • Xin Peng,
  • Peng Lin,
  • Bingxu Huang,
  • Sheng Pi,
  • Wei Wang,
  • Mingnian Wang

摘要

This study focuses on intelligent detection of groundwater seepage states on drill-and-blast tunnel faces and establishes a unified evaluation framework spanning two modalities—visible (VIS) and infrared (IR)—and five models (Faster R-CNN, SSD, DETR, YOLOv8, YOLOv11). VIS and IR images were sequentially acquired during the same inspection sessions on the same tunnel faces at four ongoing tunnel sites in western China, and, in accordance with the Railway Tunnel Design Code (TB 10003 − 2016) and on-site flow measurements, three seepage states—damp/dripping, rain-like/linear, and flowing—were annotated. All models were compared under identical data splits and unified training/inference protocols, with evaluation metrics including mAP@50, mAP@50–95, training-dynamics curves, and FPS. The comparative results show that, relative to VIS, IR provides higher overall stability and better recognition of safety-critical classes (“rain-like/linear” and “flowing”). For engineering deployment, the YOLO family offers a more favorable accuracy–speed trade-off and is therefore recommended as the primary choice.