A Novel Teacher–Student Framework for Unsupervised Domain Adaptation in Multi-Perspective Road Pavement Damage Detection

Automated detection of road pavement damage is essential for ensuring road safety and preserving the integrity of transportation infrastructure. Traditional inspection methods often suffer from high labor demands, low time efficiency, substantial costs, and potential disruptions to traffic flow. Recent advances in deep learning have significantly improved the efficiency and accuracy of pavement damage detection from remotely sensed imagery. However, these approaches face considerable challenges when dealing with multi-perspective image datasets. The inherent domain shift phenomenon severely impairs model generalization and cross-domain adaptability. To overcome these issues, we propose a novel Road Damage Domain Adaptation (RDDA) framework, specifically designed for cross-domain road pavement damage detection using remotely sensed multi-perspective images. RDDA adopts a teacher–student architecture with hierarchical feature alignment to enable Unsupervised Domain Adaptation from labeled source domains to unlabeled target domains across different viewing angles. In addition, a road augmentation module is introduced to enhance pavement texture and appearance in target domain images, thereby improving pseudo-label reliability and detection performance. Extensive experiments on diverse multi-perspective datasets, including wide-view, street-view, and drone-view imagery, demonstrate that RDDA effectively mitigates domain shift and achieves state-of-the-art performance in cross-domain settings. Notably, our framework offers a systematic integration and comprehensive empirical validation of multi-perspective domain adaptation for road pavement damage detection, providing a robust and scalable solution for intelligent infrastructure inspection.