FSA-DETR: A Frequency-Aware and Structure-Aligned DETR for Small-Object Detection in UAV Aerial Imagery

Small objects in unmanned aerial vehicle (UAV) imagery often occupy only a few pixels, making their responses vulnerable to downsampling, cluttered backgrounds, and cross-scale feature misalignment. Most detector improvements handle backbone representation, encoder context modeling, and neck fusion as separate design choices, so the frequency structure of tiny targets and the geometric consistency of multi-scale features remain underused. FSA-DETR addresses this gap with a frequency-aware and structure-aligned design built on Real-Time Detection Transformer (RT-DETR). Its central idea is to treat UAV small-object detection as a frequency–structure representation problem: high-frequency target cues are enhanced in the backbone and encoder, while local structural orientation is aligned during cross-scale fusion. The design combines a Cross-stage Spectral-Aware Multi-scale Block, a Wavelet-enhanced Intra-scale Transformer Encoder, and a Spectral Orientation-Aligned Fusion module. On VisDrone2019, FSA-DETR achieves 21.5% AP and 37.5% AP50, improving the RT-DETR baseline by 1.6% and 2.4%, respectively, while reducing the computational cost to 48.8 GFLOPs and the parameter count to 15.04 M. The small-object metric APs increases by 1.3%, and the broader size-stratified gains indicate improved sensitivity to aerial targets across multiple scales. On AI-TOD, FSA-DETR obtains 22.5% AP and 51.9% AP50, suggesting that the same frequency–structure design transfers to extremely tiny objects. Under the reported benchmark protocol, explicitly coupling frequency-sensitive perception with structure-aligned fusion improves UAV small-object detection while keeping the complete model smaller than the RT-DETR baseline used in this study; this suggests that separating spectral discrimination from geometric alignment is a useful design pattern for UAV detectors.