Aerial Small Object Detection Based on Adaptive Dual‐Branch Frequency‐Spatial Feature Fusion

Counts differ because each database indexes a different set of publications. We treat OpenAlex as the canonical count; Google Scholar is not shown (no API, and crawling it violates its ToS).

Small object detection in UAV imagery remains challenging. Existing methods still exhibit insufficient feature extraction and feature fusion capabilities, limiting their robustness under dense object distributions, complex backgrounds, blurred texture details, and scale variations. To address these issues, this paper proposes Frequency‐Spatial Decoupled YOLO (FSD‐YOLO). Specifically, a Frequency‐Spatial Decoupled Frontend (FSDF) with an asymmetric dual‐branch architecture is designed, in which the spatial branch preserves semantic representations while the frequency‐domain branch learns complementary structural information, thereby enhancing the complementary representation of spatial semantics and frequency‐domain structures. On this basis, a Saturation‐Aware Gating (SAG) mechanism adaptively regulates the contributions of spatial‐domain and frequency‐domain features according to the informativeness of channel responses, enabling adaptive residual fusion between the two domains. In addition, a Lightweight Strip Attention (LSA) module, a high‐resolution P2 detection head, and an NWD‐CIoU hybrid localization loss are incorporated to improve complex‐background suppression, fine‐grained detail preservation, and localization robustness for tiny objects, respectively. Experiments on VisDrone2019 show that FSD‐YOLOn improves mAP@50 and mAP@50:95 by 4.5 and 2.9 percentage points over the baseline, respectively, while requiring only 2.95 M parameters. End‐to‐end UAV video‐streaming experiments further demonstrate a favorable balance between detection accuracy and computational efficiency.