Attention-enhanced dual-stage deep learning for automated tunnel crack detection and quantification

Reliable detection and quantification of tunnel lining cracks are vital for metro infrastructure safety. This study proposes a real-time deep learning framework Enhanced YOLOv8 for crack localization and DeepCrack with transfer learning for geometric quantification. It forms a novel quantification-oriented dual-stage model in which attention-enhanced crack detection explicitly guides subsequent fine-grained segmentation and measurement, thereby enabling robust and real-time tunnel crack evaluation under field conditions. Validated on Nanchang Metro Line 1, the model achieves superior accuracy, with Enhanced YOLOv8 reaching an mAP@0.5 of 95.3% and recall of 91.7%, and the segmentation model maintaining an mAP@0.5 above 0.91, significantly outperforming baseline methods. Field deployment demonstrates robust adaptability, achieving mAP@[0.5:0.95] ≈ 0.70, sustaining speeds over 25 FPS with latency below 40 ms per frame, and manual workload decreases by 70%. This work advances automated tunnel inspection by delivering a high-precision, real-time, and field-validated solution with strong potential for broader infrastructure monitoring applications.