Air-YOLO:A New Small Object Detection Model of UAV

To address the challenges of small object detection in UAV aerial imagery, including low pixel resolution, lighting variations, and dense target distribution, we propose Air-YOLO. The model introduces three key innovations: 1) A Gather-andDistribute (GD) mechanism reconstructs the feature pyramid to enhance multi-scale feature fusion; 2) An EMA attentionenhanced detection head improves localization capability; 3) SoftNMS optimizes confidence adjustment for dense targets. Experiments on VisDrone and AI-TOD datasets demonstrate significant improvements in both precision and recall metrics compared to baseline models.