基于改进YOLOv8n的航拍视角小目标检测算法

打开文本图片集

关键词：无人机航拍；小目标检测；YOLOv8n；注意力机制；深度可分离卷积；多尺度上下文；特征融合；Soft-NMS中图分类号：TN911.73-34；TP391.4 文献标识码：A 文章编号：1004-373X（2026）09-0060-09

Smallobjectdetectionalgorithmbased on improved YOLOv8n foraerial photographyviewpoint

Li Jingyi¹²，Zhen Guoyong1，²，Chu Chengqun1，²，Gao Jiaqi1，²，Li Hongsheng1，2 （1.School of Instrument andElectronics，North UniversityofChina，Taiyuan O3oo51，China; 2.StateKeyLaboratoryofDynamicMeasurementTechnology，TaiyuanO3Oo51，China）

Abstract：Inview of thelowdetectionaccuracycaused bysmall scale，dense andoclusionof theobjects in the UAV aerial photographyviewpoint，aswellastheresourceconstraintsonUAV（unmannedaerialvehicle）equipment，animproved smallobjectdetectionalgorithmforUAVaerialphotographyviewpointisproposedbasedonYOLOv8n.Firstly，thereceptive field，andchannelandspatialatentionmechanismsareintegatedintotheconvolutiotoesignFAMConv（receptive-field channelspaceatentionmoduleconvolution）module，soastoimprovethemodelabilityofextractingcomplexfeatures，ndthen thedilation-wiseresidualC2f（C2f-DWR）moduleisalsointroducedtoobtainmulti-scalecontexteficientlyandenhancethe modelcapabilityofrecognizingsmallobjects.Secondly，theenhancedadaptivebidirectionalmulti-scalefeaturefusionneck network structureisdesigned，SDConv（space-to-depthonvolution）isusedtocapturefine-grainedfeatures，andtheultikernelatentionMKA（multi-kernelatention）moduleisdesignedtoenhancemulti-scalefeaturefusion.Finally，Soft-NMS （softnon-maximumsuppression）isemployedtoadjusttheoverlapping boxsuppessonstrategyandimprovethedetection accuracy of the occluded objects.Experiments onthe dataset VisDrone2O19 show that，the mean average precision mAP （ω0.5 and mAP@0.5：O.95of theimprovedmodelareenhancedby8.7%and7.6%，respectively，anditsparametersarereducedby 31.5% incomparisonwiththoseofthebenchmarkmodelYOLOv8n，whichverifytheefectivenessof theimprovedmodelinsmllobject detection for UAV aerial photography viewpoint.

Keywords：UAVaerial photography；small object detection；YOLOv8n；atention mechanism；depthwiseseparable convolution;multi-scale context; feature fusion; Soft-NMS

0 引言

随着无人机设备与目标检测算法的发展，基于无人机平台的目标检测突破了传统视角限制，在交通监控、灾害监测、救援行动、军事侦察等领域发挥着重要的作用[2-4。（剩余11969字）