WS-RTDETR: a wavelet-guided RT-DETR for small object detection in foggy UAV imagery
TL;DR
WS-RTDETR, a robust and efficient detector with a wavelet-guided stage-specific design, is proposed, which demonstrates the effectiveness of WS-RTDETR on an independent hazy UAV benchmark under the same retraining and evaluation protocol.
Abstract
Accurate small-object detection in UAV imagery is essential for low-altitude autonomous operations, but fog severely degrades image quality and makes it difficult to maintain both detection accuracy and computational efficiency. To address this problem, we propose WS-RTDETR, a robust and efficient detector with a wavelet-guided stage-specific design. Its three components serve distinct roles across the network: the frequency-domain decoupling embedding module performs branch-specific frequency processing during early downsampling by selectively normalizing low-frequency responses while retaining high-frequency structural information, the small object enhancement pyramid strengthens shallow high-resolution features for tiny-target representation, and the wavelet-guided selective encoder performs wavelet-guided dual-stage spatial refinement to emphasize informative regions and suppress redundant background responses. Experiments on the synthetic VisDrone2019-Fog dataset show that WS-RTDETR achieves 38.2% mAP50 and 22.3% mAP50:95, outperforming RT-DETR-R18 by 2.9 and 1.9 points, respectively, while improving APS by 2.6 points at 64.3 GFLOPs and 22.5 M parameters. Additional experiments on HazyDet further demonstrate the effectiveness of WS-RTDETR on an independent hazy UAV benchmark under the same retraining and evaluation protocol.