引用格式:张顺，赵倩，赵琰．基于多分支融合网络的遥感飞机检测算法［J］．电光与控制，2023，30(3):107-111．ZHANGS,ZHAOQ,ZHAOY．Aremotesensingaircraftdetectionalgorithmbasedonmulti-branchfusionnetwork［J］．ElectronicsOptics＆Control,2023,30(3):107-111．基于多分支融合网络的遥感飞机检测算法张顺，赵倩，赵琰(上海电力大学电子与信息工程学院，上海201000)摘要:针对目前遥感图像检测精度低、召回率低、实时性差等问题，提出基于GhostNet和CoT多分支残差网络(MBRNet)的遥感飞机检测算法。借鉴YOLOv4网络模型，采用MBRNet作为新的主干网络，从而减少梯度消失问题并弥补了CNN欠缺的全局特征计算能力;为了减少小目标丢失问题，同时在主干与PANet中引入多方位的特征提取与融合思路，实现在高、低特征层之间和同尺度特征层之间的信息充分互补。提出的算法在具有背景复杂、过度曝光、目标密集等场景的RSOD和LEVIR数据集上准确率达到了97．64%，召回率达到了89．11%。关键词:遥感图像;遥感飞机;多分支残差网络;YOLOv4;GhostNet中图分类号:TP753文献标志码:Adoi:10．3969/j．issn．1671－637X．2023．03．019ARemoteSensingAircraftDetectionAlgorithmBasedonMulti-BranchFusionNetworkZHANGShun,ZHAOQian,ZHAOYan(CollegeofElectronicandInformationEngineering,ShanghaiElectricPowerUniversity,Shanghai201000,China)Abstract:Aimingattheproblemsoflowdetectionaccuracy,lowrecallrate,andpoorreal-timeperformanceofremotesensingimages,aremotesensingaircraftdetectionalgorithmbasedonGhostNetandCoT(ContextualTransformer)Multi-BranchResidualNetwork(MBRNet)isproposed．LearningfromtheYOLOv4networkmodel,MBRNetisadoptedasnewbackbonenetworktoreducetheproblemofgradientdisappearanceandmakesupforthelackofglobalfeaturecalculationcapabilitiesofCNN．Inordertoreducetheproblemofsmalltargetloss,multi-directionalfeatureextractionandfusionareintroducedintothebackboneandPANet．Theideaistorealizefullcomplementationofinformationbetweenhighandlowfeaturelayersandbetweenfeaturelayersofthesamescale．Theproposedalgorithmhasanaccuracyof97．64%andarecallrateof89．11%onRSODandLEVIRdatasetsinthecircumstanceofcomplexbackground,overexposureanddensetargets．Keywords:remotesensingimage;remotesensing...
