G-YOLO v7:面向无人机航拍图像的目标检测算法  

G-YOLO v7:target detection algorithm for UAV aerial images

作　　者：陈卫彪 贾小军 朱响斌[2] 冉二飞 魏远旺 CHEN Weibiao;JIA Xiaojun;ZHU Xiangbin;RAN Erfei;WEI Yuanwang(College of Information Science and Engineering,Jiaxing University,Jiaxing,Zhejiang 314001,China;School of Computer Science and Technology,Zhejiang Normal University,Jinhua,Zhejiang 321004,China;School of Computer Science and Technology(School of Artificial Intelligence),Zhejiang Sci-Tech University,Hangzhou,Zhejiang 310018,China)

机构地区：[1]嘉兴大学信息科学与工程学院,浙江嘉兴314001 [2]浙江师范大学计算机科学与技术学院,浙江金华321004 [3]浙江理工大学计算机科学与技术学院(人工智能学院),浙江杭州310018

出　　处：《光电子.激光》2025年第2期146-157,共12页Journal of Optoelectronics·Laser

基　　金：浙江省公益技术应用研究计划项目(LGG20F010010)资助项目。

摘　　要：针对传统无人机(unmanned aerial vehicle,UAV)目标检测算法存在漏检率高、检测成功率低、模型体积大等问题,提出一种新的基于GhostNet和注意力机制的大检测头网络结构的目标检测方法G-YOLO v7(GhostNet YOLO v7)。该技术在YOLO v7-tiny的基础上增加一个大尺寸160×160目标检测头以提升小目标检测能力,同时对网络进行轻量化处理。删除原有20×20的小检测头及其卷积结构,新增GhostNet卷积模块,以减少网络的参数量,降低模型体积,同时修改损失函数为WIoU(wise intersection over union),增加PCBAM(parallel convolutional block attention module)注意力模块以提升检测精度。实验结果表明,基于G-YOLO v7网络结构的目标检测的mAP@0.5为42.3%,较YOLO v7-tiny提升5.2%,较YOLO v8n提升7.4%。G-YOLO v7的参数量和模型体积仅为YOLO v7-tiny的33.9%和37.9%,YOLO v 8n的64%和75.6%,能够有效地应用于无人机航拍图像目标检测。A new target detection method GhostNet YOLO v7(G-YOLO v7)based on GhostNet and attention mechanism with large detection head network structure is proposed to solve the problems of high missed detection rate,low detection success rate and large model volume of traditional unmanned aerial vehicle(UAV)target detection algorithm.This technology adds a large 160×160target detection head on the basis of YOLO v7-tiny to improve the small target detection ability,and lightweight processing is performed on the network.The original 20×20minimum detection head and its convolution structure are deleted,and GhostNet convolution module is added to reduce the number of network parameters and model volume.At the same time,the loss function is modified to wise intersection over union(WIoU),and parallel convolutional block attention module(PCBAM)is added to improve the detection accuracy.The experimental results show that the mAP@0.5of target detection based on GYOLO v7network structure is 42.3%,which is 5.2%higher than that of YOLO v7-tiny,7.4%higher than that of YOLO v8n.The parameter quantity and model volume of G-YOLO v7are only 33.9%and 37.9%of YOLO v7-tiny respectively,64%and 75.6%of YOLO v8nrespectively,which can be effectively applied to unmanned aerial vehicle aerial image target detection.

关 键 词：G-YOLO v7 目标检测 GhostNet YOLO v7 无人机图像(UAV) 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]