Freq-DETR: Frequency-aware transformer for real-time small object detection in unmanned aerial vehicle imagery

Recent advancements in unmanned aerial vehicle (UAV) and remote sensing technologies have propelled UAV object detection to the forefront of computer vision research. Despite significant progress in deep learning-based detection algorithms, critical challenges persist in small object detection, including high-frequency information loss, inadequate multiscale feature representation, etc. To address these limitations, this paper proposes Freq-DETR, a frequency-aware real-time transformer detection framework leveraging frequency domain analysis to enhance edge detail preservation and global contextual modeling through three novel innovations. First, the frequency-enhanced convolution module (FECM) synergistically integrates spatial and frequency features via dual-branch processing; Second, the decoupled intra-feature scale interaction module (DSC-Clo block) facilitates the integration of high-frequency local and low-frequency global information; Finally, the attention-guided selective feature pyramid network (AGS-FPN) employs context-aware attention for high-level screening feature fusion. Extensive evaluations on the VisDrone2019 benchmark demonstrate that Freq-DETR outperforms the baseline RT-DETR by 4.9 % $map@50$ gain while maintaining computational efficiency. There are also remarkable improvements on both UAVDT and HIT-UAV datasets. Ablation investigations and visual interpretability analyses further confirm the complementary benefits of its frequency-domain components and the framework’s robustness in complex aerial scenarios.