Bridging cross-domain and cross-resolution gaps for UAV-based pavement crack segmentation

Abstract

The acquisition of pavement distress images using UAVs presents unique challenges compared to ground-based methods due to differences in camera configurations, flight parameters, and lighting conditions. These factors introduce domain shifts that undermine the generalizability of segmentation models. To address these limitations, an interactive segmentation model, CDCR-ISeg, is proposed to bridge the gap between industrial requirements and existing methodologies. A dedicated dataset comprising 1500 pixel-wise annotated UAV images (UAV-CrackX4, X8, X16) was constructed, capturing various zoom levels and domain conditions to support the model's development. CDCR-ISeg incorporates super-resolution and domain adaptation techniques to enhance model generalization while reducing annotation efforts. Additionally, a vector map is introduced to improve boundary detection by embedding positive and negative clicks with reversed vector map directions. This approach effectively enables high-precision detection of pavement distress under diverse UAV parameter settings, addressing the critical challenges of adaptability and scalability in UAV-based pavement inspection.