Scholarlink

<p>The subgrade serves as the foundation for road engineering. However, subjected to cyclic loading, anthropogenic, and natural factors, it is vulnerable to a range of defects, including loosening and voids, which seriously deteriorate road performance and safety. The efficacy of traditional radar detection methods is often hampered by significant limitations in efficiency and accuracy, rendering them insufficient for the evolving needs of modern road maintenance. As a result, establishing the relationship between different subgrade defects types and the applicability of intelligent recognition algorithms has risen to the forefront as a pivotal and complex research problem attracting considerable interest. Based on field radar detection and forward modeling results, this paper established an image database of subgrade defects. It completed subgrade defect image detection experiments using three object detection networks: YOLO v3, YOLO v5, and YOLO v8. The study analyzed the training iteration variation patterns of the Loss function and mean Average Precision (mAP) and their prediction accuracy. It revealed the advantages of different image recognition algorithms for the detection effect of various subgrade defects and clarified the applicability relationship between subgrade defect types and image recognition algorithms. (1) YOLOv8 consistently achieves the highest detection accuracy and fastest convergence, particularly excelling in void identification; (2) YOLOv5 demonstrates stable and balanced performance across all three defect types—loosening, voids, and underground structures—making it a robust general-purpose solution; and (3) YOLOv3 is effective for detecting loosening and underground structures but shows limited capability for voids, especially under insufficient training. Notably, with adequate training (300 epochs), all three models attain high and comparable accuracy (&gt;0.81) for each defect category. These findings establish a practical applicability framework linking subgrade defect characteristics to algorithm selection, thereby enhancing the precision and reliability of intelligent road inspection systems.</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

Study on the Applicability of Intelligent Recognition Algorithms to Different Types of Subgrade Defects

  • 期刊论文

  • 发表日期：

    2026年1月22日

查看全文
  • Yu-xuan Han,
  • Yan-li Qi,
  • Yi Zhang,
  • Yi-guo Xue,
  • Jiang Zhao,
  • Fan-meng Kong,
  • Sheng-yong Dong,
  • Gang Li,
  • Xin-shun Li,
  • Ao-xiong Shu,
  • Hao-ran Zhu

摘要

The subgrade serves as the foundation for road engineering. However, subjected to cyclic loading, anthropogenic, and natural factors, it is vulnerable to a range of defects, including loosening and voids, which seriously deteriorate road performance and safety. The efficacy of traditional radar detection methods is often hampered by significant limitations in efficiency and accuracy, rendering them insufficient for the evolving needs of modern road maintenance. As a result, establishing the relationship between different subgrade defects types and the applicability of intelligent recognition algorithms has risen to the forefront as a pivotal and complex research problem attracting considerable interest. Based on field radar detection and forward modeling results, this paper established an image database of subgrade defects. It completed subgrade defect image detection experiments using three object detection networks: YOLO v3, YOLO v5, and YOLO v8. The study analyzed the training iteration variation patterns of the Loss function and mean Average Precision (mAP) and their prediction accuracy. It revealed the advantages of different image recognition algorithms for the detection effect of various subgrade defects and clarified the applicability relationship between subgrade defect types and image recognition algorithms. (1) YOLOv8 consistently achieves the highest detection accuracy and fastest convergence, particularly excelling in void identification; (2) YOLOv5 demonstrates stable and balanced performance across all three defect types—loosening, voids, and underground structures—making it a robust general-purpose solution; and (3) YOLOv3 is effective for detecting loosening and underground structures but shows limited capability for voids, especially under insufficient training. Notably, with adequate training (300 epochs), all three models attain high and comparable accuracy (>0.81) for each defect category. These findings establish a practical applicability framework linking subgrade defect characteristics to algorithm selection, thereby enhancing the precision and reliability of intelligent road inspection systems.