基于改进YOLOv5的轻量化无人机检测算法

Lightweight UAV Detection Algorithm Based on Improved YOLOv5

针对现有的无人机检测算法无法同时兼顾检测速度及检测精度的问题,本文提出了一种基于YOLOv5s(You Only Look Once)的轻量化无人机检测算法TDRD-YOLO(Tiny Drone Real-time Detection-YOLO).该算法首先以YOLOv5s的多尺度融合层和输出检测层分别作为颈部网络和头部网络,引入MobileNetv3轻量化网络对原骨干网络进行重构,并将骨干网络后的通道在原YOLOv5s的基础上进行压缩,减小网络模型大小;其次,将骨干网络中Bneck模块的注意力机制由SE修改为(Convolutional Block Attention Module,CBAM)并在颈部网络引入CBAM,使网络模型更加关注目标特征;最后修改颈部网络的激活函数为h-swish,进一步提高模型精度.实验结果表明:本文提出的TDRD-YOLO算法平均检测精度达到96.8%,与YOLOv5s相比,参数量减小到原来的1/11,检测速度提升1.5倍,模型大小压缩到原来的1/8.5.实验验证了本文算法可在大幅降低模型大小、提升检测速度的同时保持良好的检测性能.

Aiming at the problem that the existing UAV detection algorithms cannot simultaneously take into account detection speed and accuracy,a lightweight UAV detection algorithm,i.e.,Tiny Drone Real-time Detection-YOLO(TDRD-YOLO)based on YOLOv5s,is proposed in this paper.Firstly,the multi-scale fusion layer and output detection layer of YOLOv5s are used as the neck network and head network,respectively.MobileNetv3 lightweight network is introduced to reconstruct the original backbone network,and the channel behind the backbone network is compressed on the basis of the original YOLOv5s to reduce the size of the network model.Secondly,the attention mechanism of the Bneck module in the backbone network is modified from SE to CBAM(Convolutional Block Attention Module),and the CBAM is introduced in the neck network to make the network model pay more attention to the target features.Finally,the activation function of the neck network is modified as h-swish to further improve the accuracy of the model.Experimental results show that the average detection accuracy of the TDRD-YOLO algorithm proposed reaches 96.8%.Compared with YOLOv5s,the number of parameters is reduced by 11 times,the detection speed increases by 1.5 times,and the model size is reduced by 8.5 times.Experiments show that the proposed algorithm can greatly reduce the model size and improve the detection speed while maintaining good detection performance.