UAV-YOLO: A Resolution-Guided Feature Fusion and Attention Network for Small Object Detection in Aerial Imagery
Abstract
Small and densely distributed objects in unmanned aerial vehicle (UAV) imagery are difficult to detect because deep feature extraction progressively removes the spatial evidence needed to separate and localize neighboring targets. This paper presents UAV-YOLO, a resolution-guided multi-scale detector that couples spatial-detail preservation with adaptive feature refinement. First, a stride-4 P2 pathway extends both the top-down and bottom-up necks, providing a high-resolution prediction surface for fine object boundaries. Second, Bidirectional Feature Pyramid Network (BiFPN)-style weighted bidirectional fusion learns the relative contribution of each scale and prevents shallow detail from being diluted by coarse semantic features. Third, Efficient Multi-Scale Attention (EMA) captures grouped cross-spatial dependencies to strengthen discriminative responses with limited parameter overhead. The resulting P2–P5 architecture forms a continuous information pathway from fine spatial evidence to deep semantic context. A complete five-seed factorial study characterizes the effects of the architectural factors and identifies resolution preservation as the principal performance mechanism. On a deterministic held-out subset of VisDrone2019, UAV-YOLO achieves <inline-formula> <tex-math notation="LaTeX">$49.19~\pm ~0.15$ </tex-math></inline-formula>% mAP50 and <inline-formula> <tex-math notation="LaTeX">$30.22~\pm ~0.06$ </tex-math></inline-formula>% mAP50-95, exceeding YOLO11s by 2.97 and 2.35 percentage points, respectively. COCO-style evaluation further increases small-object AP from <inline-formula> <tex-math notation="LaTeX">$11.84~\pm ~0.22$ </tex-math></inline-formula>% to <inline-formula> <tex-math notation="LaTeX">$15.06~\pm ~0.16$ </tex-math></inline-formula>% and small-object recall from <inline-formula> <tex-math notation="LaTeX">$19.59~\pm ~0.35$ </tex-math></inline-formula>% to <inline-formula> <tex-math notation="LaTeX">$25.33~\pm ~0.14$ </tex-math></inline-formula>%, without a statistically detectable reduction in large-object AP. These results demonstrate that resolution-guided feature refinement provides an effective and compact 9.70M-parameter solution for dense aerial small-object detection.