DSFSA-RTDETR: A UAV Image Small Object Detection Network Based on Spatial-Frequency Dual-Domain Collaboration

Abstract

To address the challenges of complex backgrounds, densely packed small objects, drastic scale variations, and feature submergence in UAV aerial scenes, we propose DSFSA-RTDETR, a novel small object detection network leveraging dual-domain spatial-frequency collaboration. First, we design a dual-domain collaborative feature extraction network, termed DSF-CSPNet, which effectively integrates local spatial details with global frequency-domain information, thereby enhancing the high-frequency features of small objects while suppressing low-frequency background noise. Second, we design a Small Object Feature Enhancement Module (SFEM) to capture complementary foreground and background cues via bidirectional context modeling. This joint attention to both the target and its surroundings improves the model’s discriminative ability for small objects. Finally, we introduce a Layer-wise Adaptive Spatial Feature Fusion (LWASFF) module. Instead of relying on fixed-weight fusion strategies, LWASFF adaptively learns multi-scale feature fusion weights for each spatial location, effectively mitigating semantic misalignment among features and cross-layer gradient conflicts. Experimental results on the VisDrone and HIT-UAV datasets demonstrate that the proposed model achieves mAP50 scores of 40.2% and 80.6%, respectively, representing improvements of 3.6% and 1.5% over the baseline, while reducing the number of parameters to 15.84M. Compared with other detectors achieving comparable performance, the proposed model achieves competitive detection performance, particularly improving localization accuracy measured by mAP50:95 and maintaining strong detection capability for small objects in complex UAV scenarios.