Task-Specific YOLO Optimization for Railway Tunnel Cracks and Water Leakage: Benchmarking and Lightweight Enhancement

The safe operation of railway systems necessitates efficient and automated inspection of tunnel defects. While deep learning offers solutions, a clear pathway for selecting and optimizing the latest object detectors for distinct defects under strict speed constraints is lacking. This paper presents a two-stage, task-specific framework for high-speed tunnel defect detection. First, this study conducts a comprehensive comparative analysis of state-of-the-art YOLO models (YOLOv5s, YOLOv8s, YOLOv10s, YOLOv11s) on self-constructed datasets. This systematic comparison identifies YOLOv5s as the optimal model for crack detection, achieving an mAP@0.5 of 0.939 at 77.5 FPS, sufficient for inspection at 50 km/h. Subsequently, for the more complex water leakage detection task, this study proposes a novel lightweight enhancement to the most accurate model, YOLOv11s, by integrating the MobileNetV4 backbone and the Wise-IoU loss function. This optimization reduces computational load by 46.0% and increases speed by 17.4% to 33.97 FPS, theoretically supporting speeds up to 58.6 km/h. The main contributions of this work are twofold. First, this study conducts a systematic comparative analysis of YOLO series (v5 to v11) for distinct tunnel defect types (linear cracks vs. irregular water leakage), providing a clear selection guideline under strict speed constraints. Second, it introduces a novel, task-specific lightweight optimization paradigm, demonstrating that a one-model-fits-all approach is suboptimal for complex inspection tasks. Our study not only provides a practical solution but also establishes a valuable benchmark and optimization paradigm for real-time defect detection in tunnel engineering.