RSF-Detr：空频增强与上下文重构的路面损伤检测

Key Points

• To address low detection precision and high omission rates in pavement damage detection by proposing an improved RT-DETR model.
• Developed a dual-domain feature enhancement module (FreSCal) for better edge and target information extraction.
• Introduced a context-guided spatial feature reconstruction pyramid network (RSDFPN) for enhanced multi-scale feature fusion.
• Implemented dynamic grouped convolution shuffling with Transformer for efficient global modeling.
• Achieved significant improvements on RDD2022 and UAV-PDD2023 datasets.
• Boosted mAP@0.5 by 1.9% and 3.7% compared to baseline methods, respectively.