Scholarlink

<p>This study proposed a method for water leakage detection in tunnels using infrared vision and deep learning, which can improve stability, accuracy, and interpretability under unstable lighting conditions. Three datasets were constructed, comprising RGB images under normal illumination conditions (N-RGB), infrared images under low illumination conditions (L-IR), and infrared images under normal illumination conditions (N-IR). The proposed method integrated a Global Attention Upsample module into a U-shaped network, optimized skip connections, and cross-scale information flow to improve feature transmission across representations. Additionally, EfficientNet enhanced feature extraction capabilities as a transfer learning-based backbone. Score-Weighted Class Activation Mapping was applied to interpret the model’s inference mechanism. Experimental results showed that infrared detection under low illumination and normal illumination conditions achieved MIOU improvements of over 3.6% and 6.5%, respectively, compared to RGB detection under normal illumination conditions. The network achieved peak MIOU scores of 75.6, 79.4, and 84.5% on the N-RGB, L-IR, and N-IR datasets while maintaining a compact model size and fast inference speed. Furthermore, visual explanations revealed that enhancing low-level feature extraction was critical for encoder optimization. GAU significantly improved feature connections, and a deeper decoder stabilized feature extraction, resulting in a more stable and interpretable training process.</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

Robust Leakage Detection in Tunnels with Unstable Illumination Using an Explainable Deep Learning Approach

  • 期刊论文

  • 发表日期：

    2025年2月19日

查看全文
  • Xiangchen Yao,
  • Shuqi Ma,
  • Zhaoyuan Zhang,
  • Yuanzhen Xu,
  • Ziyi Bai,
  • Bo Li

摘要

This study proposed a method for water leakage detection in tunnels using infrared vision and deep learning, which can improve stability, accuracy, and interpretability under unstable lighting conditions. Three datasets were constructed, comprising RGB images under normal illumination conditions (N-RGB), infrared images under low illumination conditions (L-IR), and infrared images under normal illumination conditions (N-IR). The proposed method integrated a Global Attention Upsample module into a U-shaped network, optimized skip connections, and cross-scale information flow to improve feature transmission across representations. Additionally, EfficientNet enhanced feature extraction capabilities as a transfer learning-based backbone. Score-Weighted Class Activation Mapping was applied to interpret the model’s inference mechanism. Experimental results showed that infrared detection under low illumination and normal illumination conditions achieved MIOU improvements of over 3.6% and 6.5%, respectively, compared to RGB detection under normal illumination conditions. The network achieved peak MIOU scores of 75.6, 79.4, and 84.5% on the N-RGB, L-IR, and N-IR datasets while maintaining a compact model size and fast inference speed. Furthermore, visual explanations revealed that enhancing low-level feature extraction was critical for encoder optimization. GAU significantly improved feature connections, and a deeper decoder stabilized feature extraction, resulting in a more stable and interpretable training process.