Automated defect segmentation and quantification in concrete structures via unmanned aerial vehicle-based lightweight deep learning

Large-scale water-related concrete structures, such as dams, inevitably develop defects over time. Traditional manual inspections are inefficient, hazardous, and prone to high false detection rates. Unmanned aerial vehicles (UAVs) equipped with high-resolution cameras provide a safer and more efficient alternative. Although deep learning (DL) has advanced defect detection from UAV imagery, many existing approaches are not specifically designed for UAV-based inspection scenarios, where both inference efficiency and deployment constraints must be carefully considered. In addition, relatively few methods incorporate three-dimensional (3D) reconstruction for spatial localization and dimensional quantification of concrete defects, limiting the applicability in scenarios requiring precise structural assessment. To overcome current limitations, this study presents an automated framework for defect detection and quantitative assessment in large-scale concrete structures by integrating UAV photogrammetry, multi-view 3D reconstruction, and DL techniques. A lightweight defect segmentation method is developed by embedding an enhanced shifted window multi-head self-attention module into a streamlined U-Net architecture, effectively capturing both fine-grained local details and broader contextual cues. The improved attention mechanism enables efficient inter-window communication with minimal computational overhead, enhancing the network's ability to detect small and fragmented defects. To further reduce model complexity and improve deployment efficiency, knowledge distillation is applied during training, allowing the student model to maintain high segmentation accuracy with reduced computational cost. In parallel, a multi-view stereo reconstruction approach is employed to generate accurate 3D point clouds of the inspected structures. Defect locations and dimensions are then quantitatively evaluated through reverse mapping and photogrammetric analysis. The proposed framework is validated through case studies on concrete beams and high-arch dams. Experimental results demonstrate that the enhanced lightweight U-Net achieves accurate segmentation, while the 3D reconstruction enables defect-precise spatial localization and millimeter-level measurement. These findings highlight the potential of combining UAV imaging, DL, and 3D reconstruction for efficient and reliable inspection of large concrete structures.