Abstract: Aiming at the problems of various sizes, inadequate tiny target detection and dense overlapping of detection frames in air-to-ground detection of UAV, an improved YOLOv11 target detection network is proposed. The network replaces the C3K2 structure of YOLOv11 network with C3K2_D structure, which enhances the network's ability to extract irregular objects. Furthermore, the YOLOv11 network incorporates the bilinear self-attention mechanism. By combining local and non-local attention, the bilinear self-attention mechanism improves the backbone network's feature extraction capabilities. To address the issue of dense overlapping of detection frames, the MPDCIoU(Maximize position dependent combination Intersection over Union, MPDCIoU) loss function is designed to improve the regression accuracy of the bounding box. In addition, the AFPN(Asymptotic Feature Pyramid Network, AFPN) small target detection head is designed to improve the algorithm's detection performance for small targets. Experimental results show that the improved YOLOv11 network reaches 38.91% and 30.67% in mAP@0.5 and mAP@0.95 indicators, respectively, which are 1.79 percentage points and 1.76 percentage points higher than the YOLOv11 network. At the same time, the algorithm's frame rate reaches 124.6 frame/s, satisfying the real-time detection requirements. The improved YOLOv11 network presents advantages in terms of detection speed and accuracy when compared to representative target detection networks like RetinaNet, YOLOv7, and YOLOv8. By comparing the results of ablation experiments, it is found that the C3K2_D structure, the global nonlinear attention module, the MPDCIoU loss function, and the AFPN detection head can all successfully enhance the network's detection performance, the effectiveness of these four improved parts is demonstrated.

Key words: UAV target detection, YOLOv11, bilinear self attention, MPDCIoU loss function, AFPN detection head

摘要： 针对无人机空对地检测中存在的尺度不一、小目标检测效果不佳及检测框密集重叠等问题，提出了一种改进YOLOv11目标检测网络。该网络将YOLOv11网络的C3K2结构替换为C3K2_D结构，增强了网络对不规则物体的提取能力，另外在Yolov11网络中引入双线性自注意力机制，双线性自注意力机制通过融合局部注意力和非局部注意力，增强了主干网络的特征提取能力；针对检测框密集重叠的问题，设计了MPDCIoU(Maximize position dependent combination Intersection over Union, MPDCIoU)损失函数以提高边界框的回归精度；此外，还设计了AFPN(Asymptotic Feature Pyramid Network, AFPN)小目标检测头，以提升算法对小目标的检测性能。实验结果表明，改进YOLOv11网络在mAP@0.5和mAP@0.95指标上分别达到38.91%和30.67%，相比YOLOv11网络提高了1.79个百分点和1.76个百分点。同时，该算法的帧率达到124.6frame/s，满足了实时检测的要求。与RetinaNet、YOLOv7、YOLOv8等具有代表性的目标检测网络相比，改进YOLOv11网络在检测精度和速度上均表现出优势。通过消融实验结果对比，发现C3K2_D结构、全局非线性注意力模块、MPDCIoU损失函数以及AFPN检测头均能有效提升网络的检测性能，验证了这四个改进部分的有效性。

关键词: 无人机目标检测, YOLOv11, 双线性自注意力, MPDCIoU损失函数, AFPN检测头