DFS-DETR: a UAV object detection framework based on frequency-domain enhancement and state space modeling

Abstract Unmanned aerial vehicle object detection (UAV-OD) has been widely applied across various domains. However, UAV imagery is typically characterized by small object scales, high object density, and complex backgrounds, all of which pose significant challenges for accurate detection. Most existing OD algorithms rely on manually designed components, limiting their ability to adapt to the unique characteristics of UAV imagery. To address these issues, we propose DFS-DETR, an end-to-end UAV small OD framework built upon RT-DETR. First, we design a lightweight backbone network, termed C2f-WDMamba, that enhances small object detail extraction and long-range dependency modeling through the collaborative integration of dynamic dilated routing, multi-level wavelet-based high-frequency enhancement, and state space modeling. Second, we introduce the WAS-FFN module into the AIFI encoder, which leverages window-based frequency-domain modeling, channel reweighting, and learnable frequency-band weighting to enhance object details while suppressing background noise. Finally, we design a cross-scale fusion structure, named CFGN (comprising C2FM and SPGF), which improves feature alignment and small object representation by combining a center-focused enhancement module with a scale-aware gating mechanism. Experimental results on the VisDrone2019 dataset demonstrate that DFS-DETR achieves improvements of 2.6% in AP 50 − 95 and 3.8% in AP 50 over the RT-DETR-r18 baseline under identical training configurations, while reducing the parameter count by 3.34 M.