Scholarlink

<p>Cracks in structures and roads can compromise safety, reduce durability, and lead to costly failures if undetected. This study explores the application and optimization of YOLO-based deep learning models—YOLOv8-n, YOLOv11-n, and a custom variant YOLO-CrackNet—for segmentation and detection of structural cracks. YOLO-CrackNet enhances crack detection by increasing C3k2 blocks, adding C2PSA attention, and using a 312-filter head, boosting precision while remaining lightweight and real-time ready. It employs a balanced scale with depth 0.40 and width 0.30 to optimize feature learning without significantly increasing computational cost. In Phase 1, models were trained to establish baselines. YOLO-CrackNet led with a mask mAP50 of 83.60% and mAP50-95 of 54.30%, outperforming YOLOv8-n and YOLOv11-n. Phase 2 involved tuning batch size, initial learning rate (LR0), and final learning rate (LRf), where the optimized YOLO-CrackNet achieved a mask mAP50 of 86.65%, mAP50-95 of 58.19%, and a lower segmentation loss of 0.97291. This improvement outperforms baseline models, achieving + 4.32% in mAP50 and + 9.12% in mAP50-95 over YOLOv8-n, and + 4.21% in mAP50 and + 8.78% in mAP50-95 over YOLOv11-n. It also demonstrated more accurate real-world detection, stable harmonic mean of precision and recall values, and tighter precision-recall curves. The architecture, based on YOLO, outperformed other models, showcasing enhanced performance for crack detection tasks. Despite a minor increase in inference time (4 ms vs. 3 ms in lighter models), it remains much faster than recently developed architectures. Overall, YOLO-CrackNet, with tuned hyperparameters, proved highly effective for crack segmentation and suitable for real-time structural health monitoring.</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

Leveraging deep learning models for structural crack detection

  • 期刊论文

  • 发表日期：

    2025年8月9日

查看全文
  • Tanumoy Ghosh,
  • Arkaprava Gangopadhyay,
  • Lona Das

摘要

Cracks in structures and roads can compromise safety, reduce durability, and lead to costly failures if undetected. This study explores the application and optimization of YOLO-based deep learning models—YOLOv8-n, YOLOv11-n, and a custom variant YOLO-CrackNet—for segmentation and detection of structural cracks. YOLO-CrackNet enhances crack detection by increasing C3k2 blocks, adding C2PSA attention, and using a 312-filter head, boosting precision while remaining lightweight and real-time ready. It employs a balanced scale with depth 0.40 and width 0.30 to optimize feature learning without significantly increasing computational cost. In Phase 1, models were trained to establish baselines. YOLO-CrackNet led with a mask mAP50 of 83.60% and mAP50-95 of 54.30%, outperforming YOLOv8-n and YOLOv11-n. Phase 2 involved tuning batch size, initial learning rate (LR0), and final learning rate (LRf), where the optimized YOLO-CrackNet achieved a mask mAP50 of 86.65%, mAP50-95 of 58.19%, and a lower segmentation loss of 0.97291. This improvement outperforms baseline models, achieving + 4.32% in mAP50 and + 9.12% in mAP50-95 over YOLOv8-n, and + 4.21% in mAP50 and + 8.78% in mAP50-95 over YOLOv11-n. It also demonstrated more accurate real-world detection, stable harmonic mean of precision and recall values, and tighter precision-recall curves. The architecture, based on YOLO, outperformed other models, showcasing enhanced performance for crack detection tasks. Despite a minor increase in inference time (4 ms vs. 3 ms in lighter models), it remains much faster than recently developed architectures. Overall, YOLO-CrackNet, with tuned hyperparameters, proved highly effective for crack segmentation and suitable for real-time structural health monitoring.