Tunnel lining crack detection model based on improved YOLOv5

Abstract

An improved real-time tunnel lining crack detection model based on YOLOv5 is proposed. This model maintains high precision and accuracy crack detection in low-light, low-contrast and high-noise environments by introducing several effective data augmentation techniques as well as semantic context encoding (SCE) and detail preserving encoding (DPE) at the head of the network structure. It achieves 90 % precision, 91 % recall, and 92 % mAP@50. The model demonstrates better detection performance than YOLOv4-tiny, YOLOv5s, YOLOv8s, and traditional threshold segmentation method, especially in complex environments to reduce misdetection and omission. The average detection time is only 12 ms per image, demonstrating the feasibility of its real-time application. The robust and generalization performance of the model is validated in specific engineering applications, showing great potential for improving detection efficiency, cost-effectiveness, and reliability in tunnel safety assessment and disasters management.