ECA-enhanced YOLO integrated with SAHI for multi-defect inspection of tunnel linings

Traditional manual inspection of tunnel defects is inefficient for the requirement of automated inspection. This study proposes an automated defect inspection framework that integrates an Efficient Channel Attention (ECA)-enhanced YOLOv11 model with Slicing Aided Hyper Inference (SAHI) strategy (ECA-YOLO-SAHI). The proposed framework can accurately segment multiple defect types and structural components directly from large volumes of field tunnel images obtained by unmanned aerial vehicles (UAVs). A new defect dataset is constructed from multiple in-situ tunnel projects. Each image of the dataset is then annotated using artificial intelligence (AI) for efficient and precise labeling. The proposed framework achieves dual-focus segmentation of defects (cracks, leakage, damage) and structural components (segment joints, bolt holes), which enables spatial relationship analysis among defects and structures. Evaluation on three representative tunnel sections demonstrated reliable performance, achieving an Intersection over Union (IoU) of 0.7333 and a Dice coefficient of 0.8303 across all defect types, with an instance-level precision of 0.9016 and recall of 0.8786. These results confirm the robustness of the proposed framework, which offers a practical solution for automated tunnel defect diagnosis.