Vol.3,No.2张小青等：田间玉米苗期高通量动态监测方法
99
convolutions[C]//TheIEEEConferenceonComputergleshotmultiboxdetector[C]//EuropeanConference
VisionandPatternRecognition.Piscataway,NewonComputerVision.Springer,Cham,Switzerland:
York,USA:IEEE,2015:1-9.2016:21-37.
[22]HEK,ZHANGX,RENS,etal.Deepresiduallearn‐[25]RENS,HEK,GIRSHICKR,etal.FasterR-CNN:To‐
ingforimagerecognition[C]//TheIEEEComputerSo‐wardsreal-timeobjectdetectionwithregionproposal
cietyConferenceonComputerVisionandPatternRec‐networks[J].IEEETransactionsonPatternAnalysis
ognition.Washington,DC,USA:IEEEComputerSoci‐andMachineIntelligence,2016,39(6):1137-1149.
ety,2016:770-778.[26]MADECS,JINX,LUH,etal.Eardensityestimation
[23]REDMONJ,DIVVALAS,GIRSHICKR,etal.YoufromhighresolutionRGBimageryusingdeeplearning
onlylookonce:Unified,real-timeobjectdetection[C]//technique[J].AgriculturalandForestMeteorology,
TheIEEEComputerSocietyConferenceonComputer2019,264:225-234.
VisionandPatternRecognition.Piscataway,New[27]ZOUH,LUH,LIY,etal.Maizetasselsdetection:A
York,USA:IEEE,2016:779-788.benchmarkofthestateoftheart[J].PlantMethods,
[24]LIUW,ANGUELOVD,ERHAND,etal.SSD:Sin‐2020,16(1):1-15.
High-ThroughputDynamicMonitoringMethodofField
MaizeSeedling
1,2,31,21,21,2
ZHANGXiaoqing,SHAOSong,GUOXinyu,FANJiangchuan
（1.BeijingResearchCenterforInformationTechnologyinAgriculture,Beijing100097,China;2.BeijingKeyLab
ofDigitalPlant,NationalEngineeringResearchCenterforInformationTechnologyinAgriculture,Beijing100097,
China;3.CollegeofInformationTechnology,ShanghaiOceanUniversity,Shanghai201306,China）
Abstract:Atpresent,thedynamicdetectionandmonitoringofmaizeseedlingmainlyrelyonmanualobservation,whichistime-
consumingandlaborious,andonlysmallquadratscanbeselectedtoestimatetheoverallemergencesituation.Inthisresearch,
twokindsofdatasources,thehigh-time-seriesRGBimagesobtainedbytheplanthigh-throughputphenotypicplatform(HTPP)
andtheRGBimagesobtainedbytheunmannedaerialvehicle(UAV)platform,wereusedtoconstructtheimagedatasetof
maizeseedlingprocessunderdifferentlightconditions.Consideringthecomplexbackgroundandunevenilluminationinthe
fieldenvironment,aresidualunitbasedontheFasterR-CNNwasbuiltandResNet50wasusedasanewfeatureextractionnet‐
worktooptimizeFasterR-CNNtorealizethedetectionandcountingofmaizeseedlingsincomplexfieldenvironment.Then,
basedonthehightimeseriesimagedataobtainedbytheHTPP,thedynamiccontinuousmonitoringofmaizeseedlingsofdiffer‐
entvarietiesanddensitieswascarriedout,andtheseedlingdurationanduniformityofeachmaizevarietywereevaluatedand
analyzed.Theexperimentalresultsshowedthattherecognitionaccuracyoftheproposedmethodwas95.67%insunnydaysand
91.36%incloudydayswhenitwasappliedtothephenotypicplatforminthefield.WhenappliedtotheUAVplatformtomoni‐
tortheemergenceofmaize,therecognitionaccuracyofsunnyandcloudydayswas91.43%and89.77%respectively.Thedetec‐
tionaccuracyofthephenotypingplatformimagewashigher,whichcouldmeettheneedsofautomaticdetectionofmaizeemer‐
genceinactualapplicationscenarios.Inordertofurtherverifytherobustnessandgeneralizationofthemodel,HTPPwasused
toobtaintimeseriesdata,andthedynamicemergenceofmaizewasanalyzed.Theresultsshowedthatthedynamicemergence
resultsobtainedbyHTPPwereconsistentwiththemanualobservationresults,whichshowsthatthemodelproposedinthisre‐
searchisrobustandgeneralizable.
Keywords:fieldmaize;FasterR-CNN;recognition;counting;dynamicseedlingdetection
（登陆www.smartag.net.cn免费获取电子版全文） |
|
