|
|
|
| CR-YOLO: an improved network model for crack detection in tunnel linings |
| KOU Lei1, YAN Wei1, WU Zhenyu1, XUE Yu1, XIONG Qingrong2, ZHANG Yu3, WANG Lige1*
|
1. School of Qilu Transportation, Shandong University, Jinan 250002, Shandong, China;
2. School of Civil Engineering, Shandong University, Jinan 250061, Shandong, China;
3. Infrastructure Inspection Research Institute, China Academy of Railway Science Co., Ltd., Beijing 100081, China |
|
|
|
|
Abstract To address the drawbacks of cumbersome procedures and poor generalization ability in traditional crack detection methods, a CR-YOLO network for tunnel lining crack detection based on deep learning was proposed. Considering the slender morphological characteristics of cracks, the context guided block(CG Block)was incorporated into the network to fuse local and global information; by integrating channel and spatial attention mechanisms at both local and global levels, the model was enabled the model to fully capture the contextual information of crack regions. Meanwhile, a reparameterized generalized feature pyramid network(RepGFPN)module was added to improve the feature pyramid network(FPN)for object detection, which achieved more efficient fusion of multi-scale features and enhanced the capture of high-level semantic information and low-level spatial details. This method optimized the model performance under constrained computing resources and reduced inference latency without a significant increase in computational overhead. Experimental results on the self-collected dataset demonstrated that the AP50 and AP50-95 of CR-YOLO reached 90.2% and 66.4%, respectively, which represented an increase of 6.2% and 3.7% compared with the baseline model YOLOv10. The model outperformed other one-stage object detection networks(YOLOv3, YOLOv5, and YOLOv9)in terms of detection accuracy. Additionally, its inference speed reached 138.7 fps, enabling real-time detection of tunnel lining cracks.
|
|
Published: 07 July 2026
|
|
|
|
|
|
|
|
