A robust road defect detection method for UAV imagery based on synergistic attention and adaptive feature refinement

Drone-based road defect detection is essential for efficient pavement inspection, but remains challenging due to scale variation, low resolution, blurred boundaries, and interference from shadows, lane markings, vegetation, and road facilities. To overcome these limitations, this study introduces a robust method built around an enhanced model, URD-YOLOv11n. This is realized through four key innovations: First, a Spatial-Channel Synergistic Attention (SCSA) module is embedded to strengthen fine-grained feature extraction by jointly enhancing spatial and channel information. Second, a Dynamic Upsampling (DySample) module is adopted to improve the reconstruction of weak and detailed defect features from low-resolution feature maps. Third, a Large Selective Kernel (LSK) attention mechanism is incorporated to enhance adaptive multi-scale perception under complex aerial backgrounds. Finally, Wasserstein Distance Loss (WDL) is used to improve the localization stability of small and blurred defects. Extensive experiments reveal that URD-YOLOv11n attains an mAP@0.5 of 82.9% and an mAP@0.5:0.95 of 53.0%, representing notable improvements of 5.1% and 4.3%, respectively, over the baseline model. The results demonstrate that the proposed method improves the detection of fine-grained cracks, reduces missed and false detections in low-resolution UAV images, and enhances robustness to multi-scale road defects while maintaining lightweight and real-time inference performance.