基于YOLOv5的无人机航拍改进目标检测算法Dy-YOLO

Dy-YOLO:Improved UAV Image Target Detection Algorithm Base on YOLOv5

由于无人机航拍具有场景复杂多样,目标尺度变化剧烈,高速低空运动模糊等诸多特性,给目标检测带来了很大的挑战。针对无人机航拍目标检测效果不佳的问题,提出了Dy-YOLO模型,在YOLOv5的基础上引入Dynamic Head注意力,从尺度感知、空间位置、多任务3个角度探索具有注意力机制的预测头潜力;设计了C3-DCN结构和Dynamic Head注意力相互配合增强特征提取能力;此外,还使用SimOTA标签分配方式来弥补小样本的损失,并使用CARAFE(content-aware resssembly of features)上采样算子,有效增强了不同卷积特征图的融合效果。在VisDrone2019测试集上,Dy-YOLO检测的平均均值精度达到了38.2%,较基线方法YOLOv5提高了7.1%,同时与主流的检测方法相比也取得更高的检测精度。结果表明,Dy-YOLO算法对于无人机航拍检测任务具有较好的性能。

UAV aerial photography presents significant challenges,including intricate and diverse scenes,significant variations in target scale,and high-speed,low-altitude motion blur.To address poor performance of object detection in UAV aerial photography,the Dy-YOLO model is introduced.It integrates Dynamic Head attention into YOLOv5,exploring the potential of prediction heads with attention mechanisms related to scale awareness,spatial location,and multitasking.Additionally,a C3-DCN structure and Dynamic Head attention complement each other,enhancing feature extraction performance.The SimOTA label assignment compensates for small sample losses,while the CARAFE(Content-Aware ReAssembly of FEatures)upsampling operator effectively improves the fusion of convolutional feature maps.Dy-YOLO achieves an average mean accuracy of 38.2%on the VisDrone2019 test set,marking a 7.1 percentage points improvement over the baseline YOLOv5 method.It also outperforms the majority of object detection algorithms,demonstrating its effectiveness in UAV aerial photography object detection tasks.