Abstract
During the designed service life, road infrastructures will bear repeated loading conditions from vehicle weights and environmental conditions, resulting in the inevitable occurrence of road distresses including cracks, potholes, etc. The traditional inspection methods by transportation engineers are normally costly and labor-intensive. In recent years, artificial intelligence (AI)-based road distress detection methods have been widely used as convenient and automated approaches, while the AI-based methods heavily depend on a large amount of high-quality images, limiting the real engineering applications. To address the issues, this study introduces RoadDiffBox, a novel framework employing controlled image generation and semi-supervised learning. The framework addresses dataset imbalances through class control and accelerates image generation by utilizing the denoising diffusion implicit model's reverse process sampling method, while employing knowledge distillation techniques optimized for resource-constrained mobile devices. It generates diverse and high-quality road distress images with automatic bounding box annotations, substantially reducing manual labeling requirements. Test results show that RoadDiffBox demonstrates strong generalizability across geographic regions (Germany, China, and India) and shows cross-domain potential in medical imaging applications. Performance evaluations demonstrate RoadDiffBox's effectiveness, with classification models achieving an F1-score of 0.95 and detection models reaching a mean average precision (mAP@50) of 0.95 and an F1-score of 0.91 in controlled settings, while maintaining robust performance (an F1-score of 0.86 and a mAP@50 of 0.91) during on-site testing in real-world conditions. On server-class hardware, the model achieves generation times as low as 0.18 s per image. It is discovered that RoadDiffBox can serve as a scalable and efficient solution for real-time road maintenance with limited datasets.