A focusing distribution pyramid and task-aligned network for lightweight tunnel crack detection

Rapid and reliable detection of tunnel lining cracks is critical for safe tunnel operation. Current models still face challenges in complex tunnel environments, including limited accuracy, high computational cost, and insufficient compactness. This study develops a lightweight crack-oriented deep learning detection framework based on the You Only Look Once version 11n (YOLOv11n) architecture, named Tunnel Crack Detection You Only Look Once lite (TCD-YOLO-lite), with dedicated feature alignment and task-aware localization mechanisms. The model integrates a Focusing Distribution Pyramid Network (FDPN) and a Task-Align Dynamic Detection Head (TADDH). FDPN enhances weak crack representations under low-contrast conditions by adaptively adjusting spatial-channel features and redistributing multi-scale information. TADDH improves classification-localization consistency through shared feature extraction, deformable spatial alignment, and dynamic feature reweighting, leading to more stable detection of slender and irregular cracks. A composite regression loss based on Complete Intersection over Union (CIoU) and Normalized Wasserstein Distance is applied to improve localization robustness. A layer-adaptive magnitude-based pruning strategy further reduces model complexity for edge deployment. Compared with YOLOv11n, TCD-YOLO-lite improves precision, recall, Average Precision at Intersection over Union threshold 0.5 (AP@0.5), and Average Precision at thresholds from 0.5 to 0.95 (AP@0.5–0.95) by 10.3, 12.4, 9.5, and 12.4 percentage points, respectively. After pruning, TCD-YOLO-lite contains only 1.13 million parameters, reducing the parameter count by approximately 56% while maintaining high detection accuracy. When deployed on an edge platform, TCD-YOLO-lite achieves real-time performance exceeding 30 frames per second, demonstrating its potential for tunnel structural health monitoring and intelligent tunnel inspection. The code is available at https://github.com/wangjing1992/tunnel_crack_yolo .