andrecognitionbasedonDarknetnetworkandYOLOv3algorithm[J].JournalofComputerApplications,2019,39(6):16631668.ThestudyofflamerecognitionininterferenceenvironmentbasedondeeplearningmethodGAOWei,SUNYi,LIYanchao,ZHOUYonghao(StateKeyLaboratoryofFineChemicals,SchoolofChemicalEngineering,DalianUniversityofTechnology,Dalian116024,Liaoning,China)Abstract:Inrecentyears,firedetectiontechnologybasedondeeplearninghasreceivedextensiveattentionandhasbeenwidelyusedinactualworkingconditions.Thispaperstudiestheneuralnetworkalgorithmtorecognizeflamesinamulti-interferenceenvironment.ThispaperusedtheMNISTdatasettotesttherecognitionaccuracyofthreeneuralnetworks,LinearNet,GoogleNet,andResNet.Thehighestaccuraciesofthethreenetworksare98.05%,98.94%,and99.06%,theaverageaccuracieswere96.69%,98.61%,and98.80%,andtheminimumlossfunctionvaluesare0.022,0.014,and0.010,respectively.ResNethasthehighestprecision,thehighestaverageprecision,andthesmallestlossfunctionvalue.Therefore,ResNetnetworkhasthehighestperformance,whichdemonstratestherationalityofusingResNettoconstructafeaturerecognitionnetwork.TheflamefeatureextractionnetworkusedinthispaperisDarkNet53,whichintroduces5residualoperationmodules,andperformsmultipleresidualoperations,soimprovestheextractionabilityofflamefeatures.Subsequently,introducelight,sun,andflamesignsasinterferencefactorsforresearch.Atotalof200imagesoftheabovethreeinterferenceandflamesarecollectedtoformtheinterferenceflamedataset.Thisdatasetcanenablethenetworktodistinguishdifferentrecognitionobjects,distinguishtheflamefromotherinterferencefactors,andachieveaccurateflamerecognition.UsetheYOLOv3algorithmtotrainthedatasetandperformflamerecognition.Therecognitionaccuracies(m)offire,sun,fire_sign,andlightare97.07%,88.47%,100%,and90.82%,respectively,allofwhichreachahighlevel.TheaveragerecognitionaccuracyofthefourclassescanbecalculatedthroughthePRcurvegraph,whichisusedtomeasuretherecognitionperformanceofthenetworkonthedataset.Therecognitionaccuracyofthefourclassesofthenetworkreaches94.09...
