Bridge concrete spalling recognition based on generative AI synthetic weather and YOLOv11

초록

Automatic concrete spalling detection on bridge surfaces using unmanned aerial vehicles (UAV) is essential for improving inspection efficiency and operational safety. However, instance segmentation models employed in UAV-based inspections are sensitive to environment-related variations in image appearance. In practice, inspection data are usually collected under limited conditions, which reduces model reliability when applied beyond standard operating scenarios. To address this limitation, weather condition modelling strategies were examined from a data construction perspective. Three datasets derived from the same UAV inspection imagery were used: a baseline, a traditionally augmented dataset based on global statistical transformations, and a generative-AI-based dataset representing diverse inspection environments. All datasets shared identical pixel-level annotations. A unified instance segmentation model, YOLOv11 seg, was trained under identical configurations and evaluated using cross-dataset testing to assess generalization under simulated weather variability. Results demonstrate that the model trained with generative-AI-enhanced data achieves stable segmentation performance in challenging inspection environments. Meanwhile, the segmentation accuracy on the original test set remained high, indicating that improved environmental adaptability was achieved without sacrificing performance under regular inspection conditions. These findings demonstrate that generative-AI-based environment synthesis effectively reduces training data bias and enhances model reliability without additional annotation, thereby providing a practical data construction strategy for UAV-based bridge inspection in real-world applications.

키워드