关键词: 无人机（UAV）, 小目标检测, 注意力机制, 轻量化, YOLOv11

Abstract: Object detection in aerial imagery captured by unmanned aerial vehicles (UAVs) faces significant challenges, such as detecting small-scale objects, handling variations in object sizes, and addressing limited computational resources. Existing small-object detection models, often large and computationally demanding, are unsuitable for deployment on edge devices. To address these limitations, the paper proposes a lightweight model, MA-YOLOv11s (multi-attention YOLOv11s), which builds on enhancements to YOLOv11. Firstly, it introduces selective small-object detection layers to improve performance while managing computational complexity. Secondly, this paper designs two lightweight feature extraction modules, C2SCSA and C2MCA, which integrate multiple attention mechanisms to enhance feature extraction for small objects in complex backgrounds while minimizing computational cost. Finally, it replaces the traditional NMS method with Soft-NMS-SIOU, significantly improving detection accuracy and robustness in scenarios with densely overlapping objects. In experiments on the VisDrone2019 dataset, compared to the YOLOv11s model, MA-YOLOv11s achieves improvements of 8.9, 1.3, 10.9, and 9.7 percentage points in precision, recall, mAP50, and mAP50：95, respectively, with only 2.291×106 parameters and 22.4 GFLOPs of computation. The experimental results show that the improved model demonstrates exceptional small object detection performance while maintaining a compact size.

Key words: unmanned aerial vehicle(UAV), small object detection, attention mechanism, lightweight, YOLOv11