一种改进YOLOv5s小目标无人机实时检测算法

An improved YOLOv5s real-time detection algorithm forsmall target UAVs

为满足对小目标无人机检测高精度、实时性的要求,提出一种基于YOLOv(You Only Look Once version) 5s的轻量化小目标无人机实时检测改进算法。采用MobileNetV3代替原有的骨干网络,以提取小目标无人机的浅层特征信息;在网络中添加轻量级卷积注意力模块,以提取小目标无人机的深层特征信息;在原有3个检测层基础上增加小目标检测层,以提高对小目标无人机的关注度。实验结果表明,与经典YOLOv5s检测算法相比,改进算法的参数量下降30.5%,权重减少5.3 MB,实现了算法的轻量化。另外,网络精度提高3.16%,网络召回率提高1.14%,帧率增加了8。改进算法的检测精度与检测速率均得到提升,并且有效地改善了小目标无人机的漏检问题。

In order to meet the high precision and real-time requirements for small target unmanned aerial vehicle(UAV)detection,a lightweight small target UAV real-time detection algorithm based on the You Only Look Once version(YOLOv)5s is proposed.Replace the original backbone network with MobileNetV3 to extract shallow-level feature information of small target UAVs;add a lightweight convolutional attention module in the network to extract the deep-level feature information of small target UAVs;add a small target detection layer on the basis of the original three detection layers to enhance the focus on small target UAVs.Experiment results show that compared with the original YOLOv5s detection algorithm,the improved algorithm reduces the number of parameters by 30.5%,reduces the weight by 5.3MB,and achieves lightweight for the algorithm.The network precision is increased by 3.16%,the network recall rate is increased by 1.14%,and the frame rate is increased by 8.The detection precision and the detection rate of the improved algorithm have both been improved,which can effectively solve the problem of missed detection of small target UAVs.