Scholarlink

<p>Pavement crack detection plays a critical role in road maintenance, yet accurate segmentation remains challenging due to complex backgrounds, fine crack structures, and intersecting patterns. To address these issues, this study proposes a U-Net+MRF framework that integrates deep feature learning with probabilistic graphical modeling. The core innovation lies in introducing a Markov Random Field (MRF) refinement stage to enhance spatial consistency and reduce fragmented predictions produced by conventional convolutional networks. The U-Net model first generates a probability map for crack regions, which is subsequently optimized using an MRF-based graph-cut strategy to enforce neighborhood coherence. Experiments were conducted on a diverse dataset of 720 annotated pavement images. Quantitative results demonstrate that the proposed U-Net+MRF method achieves an F1-score of 90.1% and an Intersection over Union (IoU) of 82.1%, outperforming several baseline models. Notably, the integration of MRF improved the Recall from 85.6% to 86.8%, indicating superior performance in capturing fine, intersecting crack structures in complex geospatial environments. Compared to existing models such as U-Net, TopoM-CrackNet, Transformer, and MGCCrackNet, U-Net+MRF achieves the highest Precision value and F1-score. Both qualitative and quantitative analyses confirm that the integration of MRF refinement provides consistent performance gains over standalone deep learning models. The findings provide a robust and automated solution for large-scale pavement health monitoring and digital twin applications in road engineering.</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

An Integrated U-Net and Markov Random Field Approach for Pavement Crack Segmentation

  • 期刊论文

  • 发表日期：

    2026年3月23日

查看全文
  • Shi-cai Yuan,
  • Ding-zhi Liu,
  • Wei-ping Shi,
  • Pan Li,
  • Yong-qiang Li,
  • Yong Yi

摘要

Pavement crack detection plays a critical role in road maintenance, yet accurate segmentation remains challenging due to complex backgrounds, fine crack structures, and intersecting patterns. To address these issues, this study proposes a U-Net+MRF framework that integrates deep feature learning with probabilistic graphical modeling. The core innovation lies in introducing a Markov Random Field (MRF) refinement stage to enhance spatial consistency and reduce fragmented predictions produced by conventional convolutional networks. The U-Net model first generates a probability map for crack regions, which is subsequently optimized using an MRF-based graph-cut strategy to enforce neighborhood coherence. Experiments were conducted on a diverse dataset of 720 annotated pavement images. Quantitative results demonstrate that the proposed U-Net+MRF method achieves an F1-score of 90.1% and an Intersection over Union (IoU) of 82.1%, outperforming several baseline models. Notably, the integration of MRF improved the Recall from 85.6% to 86.8%, indicating superior performance in capturing fine, intersecting crack structures in complex geospatial environments. Compared to existing models such as U-Net, TopoM-CrackNet, Transformer, and MGCCrackNet, U-Net+MRF achieves the highest Precision value and F1-score. Both qualitative and quantitative analyses confirm that the integration of MRF refinement provides consistent performance gains over standalone deep learning models. The findings provide a robust and automated solution for large-scale pavement health monitoring and digital twin applications in road engineering.