Abstract: 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%.

Key words: Object detection, Lightweight network, YOLOv5s, SE channel attention mechanism, Channel pruning