基于改进轻量级卷积神经网络MobileNetV3的番茄叶片病害识别

智慧农业（中英文）SmartAgricultureVol.4,No.1
56
GuangxiUniversity(NaturalScienceEdition),2021,munication,2022,101:ID116549.
46(4):996-1007.[24]耿海波.基于U-Net模型的单声道唱声分离研究[D].
[22]AMAHANPA,VILLARICAMV,VINLUANAA.乌鲁木齐:新疆大学,2020.
TechnicalanalysisofTwitterdatainpreparationofpre‐GENGH.Researchonmonosingingsoundseparation
dictionusingmultilayerperceptronalgorithm[C]//Pro‐basedonU-Netmodel[D].Urumqi:XinjiangUniversi‐
ceedingsof20214thInternationalConferenceonDataty,2020.
ScienceandInformationTechnology(DSIT2021).[25]MABUNID,BABUSA.Highaccurateandavariant
NewYork,USA:AssociationforComputingMachin‐ofk-foldcrossvalidationtechniqueforpredictingthe
ery,2021:120-124.decisiontreeclassifieraccuracy[J].InternationalJour‐
[23]KUMARV,SINGHR,DUAY.Morphologicallydilat‐nalofInnovativeTechnologyandExploringEngineer‐
edconvolutionalneuralnetworkforhyperspectralim‐ing,2021,10(2):105-110.
ageclassification[J].SignalProcessing:ImageCom‐
IdentificationofTomatoLeafDiseasesBasedonImproved
LightweightConvolutionalNeuralNetworksMobileNetV3

ZHOUQiaoli,MALi,CAOLiying,YUHelong
（CollegeofInformationTechnology,JilinAgriculturalUniversity,Changchun130118,China）
Abstract:Timelydetectionandtreatmentoftomatodiseasescaneffectivelyimprovethequalityandyieldoftomato.Inorderto
realizethereal-timeandnon-destructivedetectionoftomatodiseases,atomatoleafdiseaseclassificationandrecognitionmeth‐
odbasedonimprovedMobileNetV3wasproposedinthisstudy.Firstly,thelightweightconvolutionalneuralnetworkMobile‐
NetV3wasusedfortransferlearningontheimagenetdataset.Thenetworkwasinitializedaccordingtotheweightofthepre
trainingmodel,soastorealizethetransferandfineadjustmentoflarge-scalesharedparametersofthemodel.Thetrainingmeth‐
odoftransferlearningcouldeffectivelyalleviatetheproblemofmodeloverfittingcausedbyinsufficientdata,realizedtheaccu‐
rateclassificationoftomatoleafdiseasesinasmallnumberofsamples,andsavedthetimecostofnetworktraining.Underthe
sameexperimentalconditions,comparedwiththethreestandarddeepconvolutionnetworkmodelsofVGG16,ResNet50andIn‐
ception-V3,theresultsshowedthattheoverallperformanceofMobileNetV3wasthebest.Next,theimpactofthechangeof
lossfunctionandthechangeofdataamplificationmodeontheidentificationoftomatoleafdiseaseswereobservedbyusing
MobileNetV3convolutionnetwork.Forthetestoflossvalue,focallossandcrossentropyfunctionwereusedforcomparison,
andforthetestofdataenhancement,conventionaldataamplificationandmixuphybridenhancementwereusedforcomparison.
Aftertesting,usingMixupenhancementmethodunderfocallossfunctioncouldimprovetherecognitionaccuracyofthemodel,
andtheaveragetestrecognitionaccuracyof10typesoftomatodiseasesunderMixuphybridenhancementandfocallossfunc‐
tionwas94.68%.Onthebasisoftransferlearning,continuetoimprovetheperformanceofMobileNetV3model,thedilated
convolutionconvolutionwithexpansionrateof2and4wasintroducedintoconvolutionlayer,1×1fullconnectionlayerafter
deepconvolutionof5×5wasconnectedtoformaperceptronstructureinconvolutionlayer,andGLUgatingmechanismactiva‐
tionfunctionwasusedtotrainthebesttomatodiseaserecognitionmodel.Theaveragetestrecognitionaccuracywasashighas
98.25%,thedatascaleofthemodelwas43.57MB,andtheaveragedetectiontimeofasingletomatodiseaseimagewasonly
0.27s,aftertenfoldcrossvalidation,therecognitionaccuracyofthemodelwas98.25%,andthetestresultswerestableandreli‐
able.Theexperimentshowedthatthisstudycouldsignificantlyimprovethedetectionefficiencyoftomatodiseasesandreduce
thetimecostofdiseaseimagedetection.
Keywords:tomatodiseaseidentification;convolutionalneuralnetworks;transferlearning;MobileNetV3;activationfunction;
identificationandclassification
（登陆www.smartag.net.cn免费获取电子版全文）