Abstract: Tunnel environments often feature complex conditions, insufficient illumination, and visual obstructions, which collectively reduce the accuracy of existing leakage-detection algorithms. To address these limitations, the authors develop an improved YOLO-B algorithm based on YOLOv8. The proposed YOLO-B model incorporates the C2f_AdditiveBlock module, which integrates local-global feature collaboration, lightweight attention, and multi-scale adaptive fusion to enhance robustness under complex and low-light conditions. A Slim-Neck module with cross-stage feature stacking is further constructed to improve detection capability for large-scale water diffusion areas while reducing computational redundancy in the neck network. Using publicly available datasets and images collected in situ by tunnel inspection vehicles, a comprehensive tunnel lining leakage dataset under realistic conditions is established. Model performance is assessed using accuracy, mAP50, mAP50－95, and GFLOPs through ablation studies and comparative analyses. The results show that: (1) relative to the baseline algorithm, YOLO-B increases mAP50 by 1.9%, mAP50－95 by 2.4%, and accuracy by 7.2%, while reducing GFLOPs to 7.3; and (2) compared with mainstream object detection algorithms such as SSD, Faster R-CNN, YOLOv5, YOLOv6, and YOLOv8, YOLO-B improves mAP50 by 6.2%, 13.2%, 3.2%, 3.7%, and 1.9%, respectively. These findings verify the effectiveness of the proposed algorithm for detecting tunnel lining leakage under complex environmental conditions.

Key words: lining water leakage, image recognition, deep learning, YOLO-B algorithm, object detection, multi-scale fusion