一种改进的基于YOLOv5s的轻量化航拍目标检测模型  

Improved Lightweight Aerial Photography Object Detection Model Based on YOLOv5s

作　　者：陈海燕[1] 毛利宏 CHEN Haiyan;MAO Lihong(School of Computer and Communication,Lanzhou University of Technology,Lanzhou 730050,China)

机构地区：[1]兰州理工大学计算机与通信学院,兰州730050

出　　处：《计算机科学》2024年第S02期465-472,共8页Computer Science

基　　金：国家自然科学基金(62161019,62061024)。

摘　　要：无人机航拍图像背景复杂、目标密集且小目标占比大,加大了目标检测的难度。基于深度学习的目标检测模型计算复杂度高,难以部署在无人机搭载的嵌入式设备上。针对此问题,提出了一种改进的基于YOLOv5s的轻量化航拍图像目标检测模型。首先将YOLOv5s主干网络的C3模块BottleNeck替换为轻量级的ShuffleNetv2网络,来降低模型的参数量和计算复杂度;其次在ShuffleNetv2网络中引入跨层信息交叉融合、SE通道注意力机制以及残差连接,来缓解卷积操作导致的特征通道数减少、网络中间层特征图的信息利用不充分问题;再次在YOLOv5s多尺度特征融合网络中引入SE通道注意力机制,来提高网络对关键特征的捕捉和提取能力;最后对改进的目标检测模型采用通道剪枝的方法使模型进一步轻量化。实验结果表明:在NWPU VHR-10数据集上,改进后的模型与YOLOv5s模型相比,目标检测的准确率和平均精度均值分别提升了3.5%,1.9%,模型的参数量和计算量降低了76%,48.7%,模型大小压缩了73.8%,检测速度提升了48%。The difficulty of target detection is increased by complex backgrounds,dense targets,and a high proportion of small objects in the unmanned aerial vehicle(UAV)aerial images.Deployment on embedded devices by drones is difficult due to the high computational complexity of the target detection model based on deep learning.Aiming at the above problems,an improved lightweight aerial image object detection model based on YOLOv5s is proposed.Firstly,the C3 module BottleNeck of the YOLOv5s backbone network is replaced with lightweight ShuffleNetv2 to reduce model parameters and computational complexity.Secondly,cross-layer information cross-fusion,SE channel attention mechanism,and residual connections are introduced in the ShuffleNetv2 network to alleviate the problem of reducing the number of feature channels caused by convolution operations and insufficient information utilization of feature maps in the middle layer of the network.Then,the SE channel attention mechanism is introduced into the YOLOv5s multi-scale feature fusion network,augmenting the network′s ability to capture and extract key features.Finally,the proposed target detection model in this paper is further lightened by channel pruning.Experimental results on the NWPU VHR-10 dataset show that,compared with the YOLOv5s model,the proposed model is increased of 3.5% in precision and 1.9% in mean average precision.The number of parameters and computational workload is reduced by 76% and 48.7%,the model size is compressed by 73.8% and detection speed improved by 48%.

关 键 词：目标检测 轻量化网络 YOLOv5s SE通道注意力机制 通道剪枝 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]