第2期2023年3月No.2Mar.2023战术导弹技术TacticalMissileTechnology基于语义分割的二阶段舰船目标检测算法研究张航，田宗浩*，李泳，彭羽茜（陆军炮兵防空兵学院高过载弹药制导控制与信息感知实验室，合肥230031）摘要：为提高图像制导弹药对大型舰船目标的毁伤效能，解决大型舰船目标关键部位漏检和定位精度差等问题，以Yolov3为基线网络，提出了基于语义分割的二阶段目标检测算法。在主体目标准确检测的基础上，利用DeepLabv3plus网络在主体目标区域进行像素级检测，确定舰船关键部位的轮廓边界，提高了舰船关键部位的检测精度和定位精度。利用Linearbottlenecks结构在低分辨率主体目标中提取特征，减少特征损失，降低模型计算量和参数量，获得了较高的检测精度和定位精度，同时提高了算法的处理速度。关键词：图像制导；舰船；目标检测；Yolov3；弹载图像；语义分割中图分类号：TN911.73文献标识码：A文章编号：1009-1300（2023）02-0137-07DOI：10.16358/j.issn.1009-1300.20220590Two-stageshipobjectdetectionalgorithmbasedonsemanticsegmentationZhangHang，TianZonghao*，LiYong，PengYuqian（LaboratoryofGuidanceControlandInformationPerceptionTechnologyofHighOverloadProjectiles，ArmyAcademyofArtilleryandAirDefense，Hefei230031，China）Abstract:Inordertoimprovethedamageefficiencyofimage-guidancemissiletothelargeships,andtosolvetheproblemsofmissinginspectionandpoorpositioningaccuracyofkeypartsoflargeships,Yolov3isusedasthebaselinenetworkandtwo-stageobjectdetectionalgorithmbasedonsemanticsegmentationisputforward.Onthebasisoftheaccuratedetectionofthemaintarget,thepixel-leveldetectioniscarriedoutinthemaintargetareabyDeepLabv3Plusnetwork,soastodeterminetheboundaryofkeypartsoftheshipandimprovethedetectionandlocationaccuracy.Inaddition,inordertoreducethelossoffeaturesandthenumberofparameters,theLinearbottlenecksmoduleisusedforextractingthefeaturesinlowresolutionimage,whichimprovesthedetectionaccuracyandtheprocessingspeedofthealgorithm.Keywords:imageguidance；ship；objectdetection；Yolov3；missile-borneimage；semanticseg‐mentation收稿日期：2022-06-30；修回日期：2022-10-30基金项目：军队“十三五”预研基金（301070103）作者简介：张航，博士研究生，主...
