Lightweight Object Detection Networks for UAV Aerial Images Based on YOLO  

作　　者：Yanshan LI Jiarong WANG Kunhua ZHANG Jiawei YI Miaomiao WEI Lirong ZHENG Weixin XIE 

机构地区：[1]ATR National Key Laboratory of Defense Technology,Shenzhen University,Shenzhen 518000,China [2]Guangdong Key Laboratory of Intelligent Information Processing,Shenzhen University,Shenzhen 518000,China

出　　处：《Chinese Journal of Electronics》2024年第4期997-1009,共13页电子学报（英文版）

基　　金：National Natural Science Foundation of China (Grant Nos. 61771319, 6207 6165, and 61871154);Shenzhen Science and Technology Project (Grant Nos. JCYJ20180507182259896 and 20200826154022001);other projects (Grant Nos. 20 20KCXTD004 and WDZC20195500201)。

摘　　要：Existing high-precision object detection algorithms for UAV(unmanned aerial vehicle) aerial images often have a large number of parameters and heavy weight, which makes it difficult to be applied to mobile devices.We propose three YOLO-based lightweight object detection networks for UAVs, named YOLO-L, YOLO-S, and YOLO-M, respectively. In YOLO-L, we adopt a deconvolution approach to explore suitable upsampling rules during training to improve the detection accuracy. The convolution-batch normalization-Si LU activation function(CBS)structure is replaced with Ghost CBS to reduce the number of parameters and weight, meanwhile Maxpool maximum pooling operation is proposed to replace the CBS structure to avoid generating parameters and weight. YOLO-S greatly reduces the weight of the network by directly introducing CSPGhost Neck residual structures, so that the parameters and weight are respectively decreased by about 15% at the expense of 2.4% mAP. And YOLO-M adopts the CSPGhost Neck residual structure and deconvolution to reduce parameters by 5.6% and weight by 5.7%, while m AP only by 1.8%. The results show that the three lightweight detection networks proposed in this paper have good performance in UAV aerial image object detection task.

关 键 词：Aerial images Object detection Deep learning You only look once Lightweight network 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术] V279[自动化与计算机技术—计算机科学与技术]