An improved real-time detection model for small targets in UAV aerial photography based on YOLOv11s
This paper addresses the challenge of small object detection in complex backgrounds under limited computational resources in unmanned aerial vehicle (UAV) aerial imagery. Based on the lightweight YOLOv11s framework, we proposes a lightweight model LMS-YOLO designed for UAV small object detection. With a view to alleviate the loss of subtle features caused by deep downsampling, we propose a local–global decoupled attention module, which splits the original global attention into local window attention and global context modules to process local details and global information respectively. Meanwhile, in order to expand the receptive field to cope with drastic scale changes, this paper proposes a multi-scale sparse dilated attention module, which can expand the receptive field without adding extra parameters through parallel processing of different dilation rates, and better capture the context around small objects. To enhance gradient sensitivity in bounding box regression and reduce missed detections of small targets, this paper proposes a loss function sensitive to small-scale objects (SS-ShapeIoU). The proposed loss function addresses low-resolution regression by penalizing edge offsets and adapting area weights. Consequently, bounding box accuracy improves significantly. Compared to the baseline YOLOv11s, LMS-YOLO improves mAP50 by 4.37% on VisDrone-DET2019, and the precision and recall are increased by 5.60% and 2.16% respectively. Additionally, our model reduces parameters by 1.91% and FLOPs by 0.83%, while sustaining an inference speed of 32.8 FPS. Evaluations on the UAVDT dataset further validate the cross-dataset adaptability of our model. Despite varying illumination and weather conditions, its mAP50 still improves by 2.49% over the baseline. In summary, LMS-YOLO achieves a good balance between accuracy and efficiency, offering strong practical viability for resource-constrained UAV edge devices.