RD-YOLO:改进YOLOv11的无人机图像小目标检测算法OA

To address the challenges of detecting small objects against complex backgrounds and the low recognition accuracy in UAV images,this paper proposed an improved model named RD-YOLO,based on the YOLOv11 architecture,to enhance the detection performance for small targets in UAV imagery.The model incorporated receptive field attention convolution(RFAConv)into the backbone network during downsampling,combined with the C3k2 module to form a C3k2-RFA module,enhancing feature extraction while mitigating spatial information loss.Additionally,a dual-oriented feature pyramid network(DoFPN)was designed as a neck network to optimize multi-scale feature fusion by balancing spatial and semantic information.The C2PSA-LSKA module,integrating large separable kernel with attention,was introduced to effectively aggregate multi-scale contextual features.Experiments on the VisDrone2019 dataset demonstrated that RD-YOLO-s achieved 46.2%in mAP@50 and 27.8%in mAP@50∶95,surpassing YOLOv11s by 6.9%and 4.5%,respectively,validating its effectiveness and advancement in small object detection for UAV aerial imagery.