改进的YOLOv4-tiny算法及其在无人机目标检测中的应用.docx

改进的YOLOv4-tiny算法及其在无人机目标检测中的应用
Abstract
Objectdetectionisakeytaskincomputervisionthatinvolvesidentifyingobjectswithinanimageorvideo.YOLOv4-tinyisanimprovedalgorithmforobjectdetectionthathasbeenappliedtothefieldofunmannedaerialvehicles(UAVs)forthedetectionoftargets.ThispaperprovidesanoverviewofYOLOv4-tinyanditsimprovedfeatures,aswellasitsapplicationinUAVtargetdetection.
Introduction
Objectdetectionisanessentialprobleminthefieldofcomputervision,andithasbeenwidelyusedinmanyapplications,includingsurveillance,autonomousdriving,androbotics.Thegoalofobjectdetectionistoidentifythepresenceandlocationofobjectswithinaninputimageorvideo.Comparedwithotherobjectdetectionmethods,YOLOv4-tinyisalightweightalgorithmthatcanachievereal-timeobjectdetectionwithfewerparameters.YOLOv4-tinyhasbeenappliedtothefieldofUAVtargetdetection,whereitiswidelyused.
ImprovedYOLOv4-tinyalgorithm
TheYOLOv4-tinyalgorithmisanimprovedversionofthepopularYOLO(YouOnlyLookOnce)algorithm.Thisalgorithmhasseveralkeyfeaturesthatenableittoachievebetterperformanceinobjectdetectiontasks.Thesefeaturesinclude:
1.Enhancementstobackbonenetwork
Thebackbonenetworkisanessentialcomponentinanyobjectdetectionalgorithm,asitisresponsibleforextractingfeaturesfrominputimages.InYOLOv4-tiny,thebackbonenetworkhasbeenimprovedtoextractmorerobustfeatures,whichhelpstoimprovetheaccuracyofobjectdetection.
2.Improveddetectionhead
ThedetectionheadisanothercrucialcomponentintheYOLOv4-tinyalgorithm.Itisresponsibleforpredictingtheboundingboxesandclassprobabilitiesofobjectswithinaninputimage.TheimproveddetectionheadinYOLOv4-tinyusesacombinationofanchor-basedandanchor-freeapproaches,whichhelpstoimprovetheaccuracyofobjectdetection.
3.Trainingstrategies
TrainingstrategiesalsoplayasignificantroleintheperformanceoftheYOLOv4-tinyalgorithm.YOLOv4-tinyusesseveraltrainingstrategies,includingself-adversarialtraining,iterativeanchorrefinement,andlabelsmoothing,whichhavebeenshowntoimprovetheaccuracyofobjectdetection.
ApplicationofYOLOv4-tinyinUAVtargetdetect