Scene classification-assisted deep learning for crack detection in asphalt pavements

Regular inspection of concrete structures is essential for maintaining their serviceability and safety. Traditional visual inspection methods face challenges due to subjective judgment, shortage of qualified engineers, and budget constraints. This research proposes an automated crack detection system using deep learning with UAV-captured visible images, addressing the challenge of varying lighting conditions on bridge pavements. A two-stage approach is implemented: first, a LeNet-based scene classification model classifies images into noisy areas, sunlit pavement areas, and shaded pavement areas; second, U-Net-based crack detection models are trained specifically for each lighting condition. To improve dataset quality and quantity, image processing techniques using L*a*b* color space are applied. Piecewise linear interpolation is used for L* values to switch contrast characteristics, while level shift is applied to a* and b* values. The scene classification achieves 94.4 % accuracy, with high precision in distinguishing between sunlit and shaded pavements. For crack detection, data augmentation through image processing significantly improves the shaded pavement model's performance, with recall increasing from 0.162 to 0.732. The highest F-measure of 0.784 is obtained when both training and test data are unified through image processing. The composite model combining scene classification and crack detection effectively reduces false detections in noisy areas and enables accurate detection across different lighting conditions. These results demonstrate that limiting training conditions to specific lighting environments and unifying dataset appearance through image processing contribute to improved crack detection accuracy in real-world bridge inspection scenarios. • Deep learning models detect cracks in UAV images under different lighting conditions. • LeNet classifies images into noisy, sunlit, and shaded pavement areas. • Image processing converts lighting conditions in pavement areas. • Unified lighting conditions in training and testing improve accuracy.