第37卷第2期2023年6月西昌学院学报（自然科学版）JournalofXichangUniversity（NaturalScienceEdition）Vol.37，No.2Jun.，2023基于改进YOLOv5的遥感图像飞机目标检测孙廨尧1，2，侯秀丽1，罗青青1（1.安徽商贸职业技术学院信智学院，安徽芜湖241000；2.芜湖市物联网智慧交通工程技术研究中心，安徽芜湖241000）摘要：针对遥感图像中飞机检测尺寸大小不一、背景复杂导致的难以识别问题，提出一种基于YOLOv5网络模型的改进方法。首先，在YOLOv5网络模型中融入Swin-Transformer模块，使网络全局建模并使全维度信息交互，以提升网络的特征提取能力；然后，对损失函数进行优化，引入SIOU损失函数以考虑真实框和预测框之间的向量角度问题。对比实验结果表明：改进后的网络模型召回率为91.2%，比改进前提升0.6个百分点；改进后平均检测精度mAP0.5为95.7%，比改进前提升0.2个百分点。改进后的YOLOv5网络模型能在一定程度上提升遥感图像中飞机目标检测性能。关键词：遥感图像；YOLOv5；Swin-Transformer；SIOU中图分类号：TP751文献标志码：A文章编号：1673⁃1891（2023）02⁃0066⁃06AircraftTargetDetectionBasedonImprovedYOLOv5forRemoteSensingImagesSUNXieyao1，2，HOUXiuli1，LUOQingqing1（1.XinzhiCollege,AnhuiVocational&TechnicalCollegeofBusiness&Trade，Wuhu，Anhui241000，China；2.WuhuInternetofThingsSmartTransportationEngineeringTechnologyResearchCenter，Wuhu，Anhui241000，China）Abstract:Aimingatthedifficultrecognitionproblemcausedbythedifferentsizesofaircraftdetectionandcomplexback⁃groundsinremotesensingimages，animprovingmethodbasedontheYOLOv5networkmodelisproposed.Firstly，theSwin-TransformermoduleisincorporatedintotheYOLOv5networkmodeltomodelthenetworkgloballyandmakethefull-dimensionalinformationinteractivetoimprovethefeatureextractioncapabilityofthenetwork；then，thelossfunc⁃tionisoptimizedandtheSIOUlossfunctionisintroducedtoconsiderthevectorangleproblembetweentherealframeandthepredictedframe.Thecomparisonexperiments'resultsshowthattherecallrateoftheimprovednetworkmodelis91.2%，whichis0.6percentagepointshigherthanthepre-improvementrate，andtheaveragedetectionprecisionmAP0.5is95.7%，whichis0.2percentagepointshigherthanthepre-improvementrate.Theim...
