智慧农业SmartAgricultureVol.1,No.2
44
Cornplantdiseaserecognitionbasedonmigrationlearning
andconvolutionalneuralnetwork
GuifenChen,ShanZhao,LiyingCao,SiweiFu,JiaxinZhou
渊130118
CornisoneofthemostimportantfoodcropsinChina,andtheoccurrenceofdiseasewillresultinseriousyield
reduction.Therefore,thediagnosisandtreatmentofcorndiseaseisanimportantlinkincornproduction.Underthebackgroundof
bigdata,massiveimagedataaregenerated.Thetraditionalimagerecognitionmethodhasalowaccuracyinidentifyingcornplant
diseases,whichisfarfrommeetingtheneeds.Withthedevelopmentofartificialintelligenceanddeeplearning,convolutional
neuralnetwork,asacommonalgorithmindeeplearning,iswidelyusedtodealwithmachinevisionproblems.Itcan
automaticallyidentifyandextractimagefeatures.However,inimageclassification,CNNstillhasproblemssuchassmallsample
size,highsamplesimilarityandlongtrainingconvergencetime.CNNhasthelimitationsofexpressionabilityandlackof
feedbackmechanism,anddataenhancementandtransferlearningcansolvethecorrespondingproblems.Therefore,thisresearch
proposedanoptimizationalgorithmforcornplantdiseaserecognitionbasedontheconvolutionneuralnetworkrecognitionmodel
combiningdataenhancementandtransferlearning.Firstly,thealgorithmpreprocessedthedatathroughthedataenhancement
methodtoexpandthedataset,soastoimprovethegeneralizationandaccuracyofthemodel.Then,theCNNmodelbasedon
transferlearningwasconstructed.TheInceptionV3modelwasadoptedthroughtransferlearningtoextracttheimage
characteristicsofthediseasewhilekeepingtheparametersunchanged.Inthisway,thetrainingprocessoftheconvolutionalneural
networkwasacceleratedandtheover-fittingdegreeofthenetworkwasreduced.Theextractedimagefeatureswereusedasinput
oftheCNNtotrainthenetwork,andfinallytherecognitionresultswereobtained.Finally,themodelwasappliedtothepicturesof
corndiseasescollectedfromthefarmlandtoaccuratelyidentifyfivekindsofcorndiseases.Identificationtestresultsshowedthat
usingdatatoenhancetheCNNoptimizationalgorithmandthemigrationstudyontheaveragerecognitionaccuracymaindiseases
ofcom(spot,southernleafblight,grayleafspot,smut,gallsmut)reached96.6%,whichcomparedwithsingleCNN,hasgreatly
improvedtheprecisionandidentificationprecisionby25.6%onaverage.Theaverageprocessingtimeofeachimagewas0.28s,
shortensnearly10timesthanasingleconvolutionneuralnetwork.Theexperimentalresultsshowthatthealgorithmismore
accurateandfasterthanthetraditionalCNN,whichprovidesanewmethodforidentificationofcornplantdiseases.
deeplearning;convolutionalneuralnetwork;transferlearning;dataenhancement;identificationofcorndisease |
|
