Chat with us, powered by LiveChat Python - Wridemy Bestessaypapers
Login

01 Feb

Posted at 17:39h in Python

python case study Python is a lagre non-venomus snake. Requirements: ACNNBasedModelforVenomousandNon-venomousSnakeClassiÞcationNagifaIlmaProgga1(B),NoortazRezoana1(B),MohammadShahadatHossain1,RaihanUlIslam2,andKarlAndersson21DepartmentofComputerScienceandEngineering,UniversityofChittagong,Chittagong,Bangladeshhossainms@cu.ac.bd2DepartmentofComputerScience,ElectricalandSpaceEngineering,Lule˚aUniversityofTechnology,Skellefte˚a,Sweden{raihan.ul.islam,karl.andersson}@ltu.seAbstract.Snakesarecurved,limbless,warmbloodedreptilesofthephylumserpents.Anycharacteristics,includingheadform,bodyshape,physicalappearance,textureofskinandeyestructure,mightbeusedtoindividuallyidentifynonvenomousandvenomoussnakes,thatarenotusualamongnon-expertspeoples.Astandardmachinelearningmethod-ologyhasalsobeenusedtocreateanautomatedcategorizationofspeciesofsnakedependentuponthephotograph,inwhichthecharacteristicsmustbemanuallyadjusted.Asaresult,aDeepconvolutionalneuralnetworkhasbeenproposedinthispapertoclassifysnakesintotwocategories:venomousandnon-venomous.Asetofdataof1766snakepicturesisusedtoimplementsevenNeuralnetworkwithourproposedmodel.Theamountofphotographsevenhasbeenincreasedbyutilizingvariousimageenhancementtechniques.Ultimately,thetransferlearn-ingmethodologyisutilizedtoboosttheidentificationprocessaccuracyevenmore.Five-foldcross-validatingforSGDoptimizershowsthattheproposedmodeliscapableofclassifyingthesnakeimageswithahighaccuracyof91.30%.WithoutCrossvalidationthemodelshows90.50%accuracy.Keywords:Snake?CNN?Dataaugmentation?Deeplearning?Transferlearning?Crossvalidation1IntroductionSnakesareectothermic,amnioticreptiles,surroundedinsepals,justlikeothersquamates.Severalsnakespecieshaveskullswithaslewofjointsthantheirreptileancestral,allowingittoswallowpredatorswiththeirextremelymaneu-verablejawsrelativelylargeunlikethereownheads.Therearetwotypesofsnakessuchasnon-venomous(non-poisonoussnake)andvenomous(poisonousN.Rezoana?equalcontribution.cSpringerNatureSwitzerlandAG2021M.Mahmudetal.(Eds.):AII2021,CCIS1435,pp.216?231,2021.https://doi.org/10.1007/978-3-030-82269-9_17 SnakeClassification217snake).VenomoussnakesaremembersofthesuborderSerpentsandareabletodevelopvenomthattheyusetoattackprey,defendthemselvesandhelpdigesttheirprey.Utilizingholloworgroovedfangs,thevenomisusuallyreleasedbyinjection,whileothervenomoussnakeslackwell-developedfangs.Non-venomoussnakes,exceptformassiveconstrictorsnakessuchastheGreenAnacondaortheBurmesePython,aregenerallybenigntohumans.Likevenomoussnakes,non-venomoussnakeshaveteeth.Snakeenvenomingisamajor,worldwidecommonhealthissuewiththegreatestprevalenceinSoutheastAsia.Fig.1.Top5countrieswiththehighestrateofsnakebitedeathsper100,000peopleAanalysisfocusedon60articleshasreportedthat363victimswithsnakebites,bothvenomous(88%)andnonvenomous(12%)werediagnosedandtreatedifnecrosisexists(15.2%)[30].Noinfectionsweredetectedinpatientsalthoughtheantibioticswerenotused.Thus,basedontheanalysis,itcanbeimpliedthatantibioticsarepresentinthesnake,consideringthefactthatverylittlerawdataisgiven.Inabilitytoidentifysnakefromthevisiblecharacteristicsisanimportantcauseofmortalityduetosnakebite.Forcenturies,snakevenom,particularlyinChineseculture,havealsobeenusedasmedicinetools.Anyoftheleadingdrugsforhighbloodpressure,cardiovasculardisease,andheartattacksusedsnakevenomasablueprint.Asaconsequence,snakevenomisoftenknownasamini-drugrepository,witheachmedicinebeingclinicallye?ective.Forexample,theFDAhaslicensedmedicinesrelyingonsnakevenom,suchasCaptoprilR(Enalapril),IntegrilinR(Eptifibatide),andAggrastatR(Tirofiban)[24].Asidefromtheseapprovedmedicines,variousothersnakevenommaterialsfornumerousthera-peuticapplicationsarepresentlyinpre-clinicalorclinicaltrials.Snakebiteenvenomingisawell-knowntropicaldiseaseexacerbatedbyaven-omoussnakeaccidentallyinjectinganextremelymodifiedpoisonintohumans.Thesnake?sfangs,thatareremodeledteethattachedtoapoisonglandsviaainletareusedtoextractthepotion.Nearly2millionresidentsinAsiaare 218N.I.Proggaetal.poisonedbysnakeseveryyear,althoughthereareapproximately435,000to580,000snakebitesannuallyinAfricathatrequiremedication.Women,childrenandfarmersinvulnerableremoteregionsinlowandmiddle-incomecountriesarea?ectedbyenvenoming.Thegreatestriskexistsinplaceswherehealthcareservicesareinadequateandmedicalfacilitiesareinsufficient.Thedetectionofsnaketypesisakeyaspectofthediagnosisprocess.Classi-ficationanddetectionofvenomousandnon-venomoussnakehasmanydynamicapplicationssuchas:Itcanhelpthehealthprofessionalindeterminingthebestanti-venomtouse.Snakevenomisoftenknownasamini-drugrepository,sobyclassifyingthevenomoussnakefromthenonvenomoussnakethevenomcanbeusedtoproducemoredrugsformankind.2RelatedWorkTheDeepConvolutionNeuralNetwork(CNN)hascarriedoutinpreviousdecadesanumberofgroundbreakingfindingsonimagerecognition,objectiden-tification,semanticclassificationetc.[21].Nowithasbecomeprominentinresearchforsomedaybecauseitcanhandleanimmensenumberofdata[4].CNNoutperformsthemachinelearningapproachintermsofefficiency.Classifi-cationandidentificationofimageswasaniconicmachinelearningissue,whichwasovercomethroughdeeplearning.Inparticular,technologiesdealingwiththevisualinformation,includingthebiggestimageclassificationdataset(ImageNet),computervisionandNLPanalysiswerequiteimpressiveandtheout-comeshavebeenobtainedisextraordinary.ThatiswhywechooseCNNforoutstudy.Thealgorithmsofdeeplearningaredesignedtoimitatetheroleofthehumanbrain[8].Inbiologicaldata[22,23]deeplearninghasbeenused.AlsoinobjectdetectionCNNhasshownoutstandingperformance[11,27,37].CNNhasalreadyshownremarkablee?ectivenessindiagnosticimagerecognition,suchasthedetectionoflungcancer[14],skincancer[28],oralcancer[18],braintumors[10],andotherdiseases.ByUsingobjectdetectionandamachinelearningmethodfocusedonMaskRegion-basedConvolutionalNeuralNetwork(MaskR-CNN)hasbeenimple-mentedforsnakespeciesclassification[7].Todiscernreptilespeciesfromimages,deeplearning,specificallyCNN,hasbeenused[5].Snakestingmarksimageswerealsousedtoidentifypoisonousandnon-poisonoussnakesusingtheCNNmethod[20].Jamesetal.[16]presentedasemiautomaticmethodinordertodi?erentiatesixdistinctorganismsbyeliminatingtaxonomicalcharacteristicsfromthepho-tos.Therewere1,299photographsinthedatasetand88photosinthelowestfrequentlevel.Theouteredgetaxonomicalcharacteristicsareleastessentialfororganismrecognitionthantheintofrontandside-viewcharacteristics,accordingtovariousfeaturedetectiontechniques.Inobjectdetectionandimageprocessing,variousmodelarchitectureshavebeenevaluated.Todiscern5specificsnakevarietiesAbdurrazaqetal.[2]used3distinctConvolutionalNeuralNetwork(CNN)frameworks.Inthisstudythey SnakeClassification219workedwithasetof415pictures.Therewere72photosrequiredforthelesscommonsnaketype.Themaximumoutcomeswereachievedbyamedium-sizedclassificationframework.TheSiamesenetworkwasusedbyAbeysingheetal.[3]tocategorizeacom-parativelylimiteddatasetthatcontains200photographsof84organismsfromtheWHO.Themethoddiscussedintheirpaperfocusesonone-shootlearning,as3to16photosperhabitatwereprovidedinthecollectionofdata.Theoutcomescollectedbytheautomaticcategorizationprocessislowerthantheexactnessofhumanidentificationaccuracy.Snakebitepoisoningisanoverlookedenvironmentaldiseaseswhichmurdersover100,000inhabitantsandslaughtersover400,000peryear[36].Snakebiteisafrequentworkplacehazardforcitizenswhomaketheirlivingincultivation,includingthoseinSouth-EastCommunitiesinasia.Anautomatedidentifica-tionofaspeciesofsnakedependingonitsphotographhasalreadybeencon-structed,aspreviouslymentioned.Asaclassifier,aseveralsupervisedmachinelearningtechniquehasbeenimplementedlikeNaiveBayes,DecisionTreeJ48,k-NearestNeighbors,orBack-PropagationNeuralNetwork.Therequirementsinthefeaturesextractionprocess,ontheotherhand,arenoteasytotraininconventionalmachinelearningtechniquesandmustbemanuallycalibrated.Asaresult,throughoutthisstudy,importantcontributionshasmadeinthefollowingareas:?ACNNbasedmodelhasbeenproposedtoclassifyvenomousandnon-venomoussnake.?Variousarchitectureshavebeencomparedintermsofperformance.?K-foldcrossvalidationhasbeenappliedonourproposedmodelwiththreedi?erentoptimizertoimprovesgeneralizationcapacity.?Thesystemcandetectbothvenomousandnon-venomoussnakesinrealtime,allowingnon-expertstorecognizesnakespecieswithgreateraccuracythanpreviouslymentionedapproaches.3DataPre-processingAtfirst,wewilldiscussaboutthedataset.Toinitiate,thedatasetandmodelcreationplanwillbediscussedindepthtoaidintheplanningoftheproposedmodel.Followingthat,we?llgothroughtheproposedmodelarchitecturesim-ulationmethodindetail,aswellasthetrainingmethodologyfordeterminingthebestparametermodifications.Ultimately,we?llusemodelingapproachestodemonstratecriticalflawsinvisualindicatorsinadditiontocreatingareportedsnakemoreidentifiable.3.1AboutDatasetThesetofdatacomesfromkaggle.comandcontainsabout1766snakeimages.Perphotowasassignedtoacategoryandwasdividedintogroupsbytherespec-tiveclasslabelssuchasnon-venomousandvenomous.Sincereformattingis 220N.I.Proggaetal.amongthemostimportantphasesofdatapreprocessing,allimagesarerefor-mattedto224×224pixels.Figure3andFig.4depictsseveralphotographsfromthebenchmarkdataset.Figures3and4demonstratethattherearenumerousdi?erencesbetweenvenomousandnon-venomoussnakesintermsofphysicalappearancessuchasheadstructure,eyeshape,skincolour,andsoon.Themen-tionedfeatureswillaidourproposedmodelinlearningthedistinctionsbetweenpoisonousandnon-poisonoussnakes.Thesetofdatahasbeensplitintotrain,validation,andtestsegmentsinanappropriateproportionFig.2.Non-venomoussnakeimageFig.3.Venomoussnakeimage3.2DataAugmentationItiswellestablishedthatevenamassivequantityofdatainthedatasetsisneededtoachieveabetterresultforaCNNmodel.ThedataaugmentationtechniquesisnecessaryforcorrectlyimplementingaCNNarchitecture.Thismethodologypreventsdatamanipulationandmaintainstheinitialreliability.Thistechniqueisoftenusedduringthetrainingprocesstoincreasetheefficiencyofthearchitecturebyfixingoverfittingproblems.Ifthedatasetislargeenough,severalfeaturescanbeextractedfromitandcomparedtounidentifieddata.However,ifthereisinsufficientdata,dataaugmentationmaybeimplementedtoboostthemodel?saccuracy[9,15,26,33].Throughapplyingaugmentationoper-ationstotrainingimages,suchasrandomrotation,shift,zoom,noise,flips,andsoon,dataaugmentationwillgeneratemultiplepictures[33].Everyparame-terhastheabilitytorepresentphotosinanumberofaspectsandcomeup SnakeClassification221withparticularfeaturesduringthetrainingphase,increasingtheframework?se?ectiveness.Sinceourdatasetissmallerinsize,weimplementedavarietyofaugmentationfunctionsonit.Fortheaugmentation,ImageDataGeneratorwasutilized.Figure5representstheprimarypictureaswellastheaugmentedphotoscreatedfromthat.Weimplementedthemodelwith80%oftheoverallofthepicturesandusedtheremaining20%tovalidatethesystemthroughoutevalua-tion.ThesettingsforimageaugmentationusedinourexperimentcanbeseeninTable1.Fig.4.DataaugmentationTable1.ImagesaugmentationsettingsAugmentationsettingRangeRotation0.2Zoom0.1Contast0.1HorizontalflipTrue4MethodologyConvolutionalneuralnetworksisinspiredbyneurologicalmechanisms.Aconvo-lutionalneuralsystemismadeupofmanylayers,includingconvolutionlayers,poolinglayers,andfullyconnectedlayers,anditusesaback-propagationalgo-rithmtoobtainfeaturestotrainthemodelproperly.Figure6depictstheoverallresearch?ssystemflowchart.4.1ModelConstructionInthisstudy,theframeworkwasimplementedusingaConvolutionalNeuralNet-work(CNN)anddataaugmentation.First,thearchitectureusesthedatasettotakethepictures.Thenpreprocessingbegins.Thensomeaugmentationparam-etersareusedtoenlargethedataset.Ultimately,theenlargedsetofdatais 222N.I.Proggaetal.Fig.5.SystemflowchartusedtoforecasttheclassbytheCNNarchitecture.Thepictureswerestan-dardisedtosomeextenttobeproperlycategorized.CNNitselfperformedthecharacteristicretrievalofthepictures.We?lldescribeourrecommendedmodelarchitectureindetailinthissection.Theproposedmethodincludesthreebasiccomponents:featureextraction,identification,andclassification.Inthebegin-ning,weincludesynchronizedlayersofconvolution,activation,andmax-poolinginoursystemdesign.Thefeaturesarethenpassedintoaflattenedlayer.Theflattenedattributesorfeaturesarethentransferredintodenselayers.Inthedenselayerdropouthasbeenusedtoavoidoverfitting.Theclassificationpro-cesswasthencompletedbytheultimatelayer,whichstatedthesoftmaxlayer.Thephotoswereprovidedtothemodelforthetrainingprocessafterthedataaugmentation.ThereisaCBr=iconvblockinthesystembuilt,accompaniedviaamaxpoolingstratumthatisthreetimesinconsonanceinsequentialformat.Maxpoolingisanefficientwayofdownsizingthetightly-scalephotos,requiringmaximalratesforeachstratum,sincemostofthefeaturesgeneratedareignoredbyutilizinga3×3filtersize.Asdiscussedduringtheprecedinganalysissuperim-posedmaxpoolingframesdonotsignificantlyincreaseoverthenonoverlappingwindows,sothemaxpoolinglayersusedduringourexperimentwas2×2withstride2.Itconsistsof3convolutionarylayers,with16filtersof3×3inthefirstconvolutionallayer,32filtersinthesecondconvolutionallayersandthethirdlayeris64layers3×3indimension.Thescaleofthekernelis2.Theproducedfeaturesarereusedbytheattachedactivationfunctionintheearlystagestoconstructanuniquefeaturemapasoutputjustaftertheconvolutionlayer.Inequation,convolutionoveranimageq(x,y)isdefinedusingafilterp(x,y).p(x,y)??q(x,y)=jy=??jky=??kw(m,n)f(x??m,y??n)(1)Togeneratethelow-levelfeaturesalllayersofconvolutionwasaccompaniedbytheReLUactivationfunction.Followingthepreviousconvolutionlayers,theReLUactivationfunctionwasusedinthehiddenlayeraswell.ReLUhasanumberofadvantages,themostnotableofwhichisitsabilitytoquicklydis-tributegradients.Asaresult,calculatingtheessentialcharacteristicsofCNNintheprovisionalmass,reducesthechancesofgradientextinction.Theactivat-ingmechanismtendstoperformcomponent-by-componentoperationsonthisgiveninputfeaturemap,sincetheresulttendstobetheidenticaldimensionas SnakeClassification223Fig.6.Modelarchitecturetheorigin.Asaresult,amongthemostcommonactivationfunctions,ReLUisbeingusedinthealllayers.ReLU[12]hasbeenincludedtodemonstratethatthearchitecturedoesnotexperiencelinearity,asshowninthecorrespondingformula:R(x)=max(0,x)(2)Thearchitectureisthenexposedintheflattenedlayerstoconvertthefeaturemap.Fromfeaturemapgeneratedbythepreviousconvolutionlayerstocompletethecategorizationtask,asingledimensionalfeaturevectorhasbeenconstructedbyflattenedlayer.Adrop-outlayerislinkedfollowingtheFullyconnectedorhiddenlayerandthesizeofthefullyconnectedlayeris128.ThisdroppinglayerwilldumptheweightsoftheFullyconnectedlayeratrandomduringtrainingtomitigateunnecessaryweightandoverfitting.ForourCNNmodel,wechoose0.2asthedropoutrangeforthedropoutlayer.Thedropoutlayer?sjobistodiscard20%ofthenodesineachFClayerinordertopreventoverfitting[29].Thehyperparameterswerefine-tunedbylayeringagainuntilfaultdidnotradicallyalter.Inaddition,therangeofthedropoutfortheproposedmodelwasdeterminedbyexperimentingforseveralconstants.Thedrop-outsrangeof0.2wasconsideredasthebestamountofthedropoutlayerbecauseittendedtopreventoverfittingratherthanothersvalue.Inordertomaximizetheoutcometheparametersarebeingadapted.Ultimately,sincetherearetwocategoriestheoutputlayercontainstwonodes.Thesoftmax[32]activationfunctionwasusedjustaftertheFClayer,asshowninthefollowingequation,toclassifysnakeimagesintonon-venomousandvenomouscategories.Figure8portraystherepresentationsofthesuggestedmodel?slayers. 224N.I.Proggaetal.Softmax(x)=eiii=0ei(3)Fig.7.Theproposedmodel?sdetailedlayerrepresentation4.2TheImplementationProcedureThecodingfortheframeworkwasproducedandimplementedinGoogleColab[6]usingthepythonprogramminglanguage.Keras[13],Tensorflow[1],NumPy[35],andMatplotlib[31]werethelibrariesutilizedinthewholestudy.Theback-endoftheframeworkwaschosenasTensorflow,andkerashasbeenutilizedtoo?eradditionalbuilt-infunctionalitysuchasactivationfunctions,optimizers,layers,andsoon.KerasAPIwasusedtoenhancethedataset.NumPyisaPythonlibraryformathematicalevaluation.Confusionmatrix,splittrainandtestfiles,modelcheckpoint,callbackmechanism,aswellasotherschematicrep-resentationlikeconfusionmatrix,lossagainstepochsgraphs,accuracyagainstepochscurves,andmanymore,areallgeneratedusingSklearn.Thematplotliblibraryisalsoneededtocreatevisualrepresentationsofthepreviouslymentioneddiagrams,suchastheconfusionmatrix.5ExperimentalEvaluationTheperformanceoftheimplementedmodeltoclassifyphotosofsnakesspecif-icallygroupedintotwogroups,non-venomousandvenomous,willbecoveredinthissection.Inaddition,wewillcompareourproposedmodeltoothertra-ditionalmodelssuchasInceptionNet,Resnet50,VGG16,VGG19,Xception,MobileNetv2andInceptionResnetV2.Theoutcomesofourmodel?skfoldcrossvalidationfordi?erentoptimizerswillalsobearticulated. SnakeClassification2255.1TuningoftheHyper-parametersSincefine-tunedhyper-parametersalsohavemajorimpactontheCNNarchi-tecture?sperformanceandtheyareareessentialbecausetheystronglyimpactthemodel?sattitude.TheAdam,Adamax,andSGDoptimizerswereutilizedtotrainIing100epochsfortheproposedmodel,withalearningrateof0.0001withthebatchsizeof32.Kfoldcrossvalidationhasalsobeenperformedforthevariousoptimizers.Asthelossfunctioncategoricalcrossentropywasbeingused,thelossofclassprobabilitycausedbythesoftmaxfunctionwasalsodeterminedbythislossfunction.Finally,calculateeachcategory?sprobabilityofoccurrence.ModelCheckpointhasbeenaddedasacallbackfunctionaswell.5.2K-FoldCrossValidationTotestourmodel,weattempt5foldcross-validation.Duetothegeneralopera-tionaloverhead,thisphaseisnormallyavoidedinCNNs.ThedatasetisdividedintothreesectionsusingK-foldscross-validation.Fold1iscomposedofpart1asatrainingset,part2asavalidationset,andpart3astestingset,whilefold2isconsistsofpart2,part1andpart3asatraining,validationandtestingsetrespectively.Thus,thefoldcontinuesuntilitreaches5folds,witheachfoldcontaininguniquetraining,validation,andtestingdatasets.TheK-foldcrossvalidationapproachaccountsfortheutilizationofvarioustrainingandtest-ingdata,whichreducesoverfittingandimprovesgeneralizationcapacity.Asaconsequence,wemaygeneralizeouroutcomesoverthedataset.5.3ResultFigure8depictstheaccuracyoftheproposedmodelascomparedtootherstan-dardConvolutionneuralnetworkslikeInceptionNet,VGG19,Resnet50,Xcep-tionNet,MobileNetv2,InceptionResnetV2andVGG16.Despiteprovidingadatasetwithalimitednumberofphotographs,thedesignedmodelproducesahighcategorizationaccuracycpmparedtootherCNNmodels.AsshownintheFig.9InceptionNet,VGG19,Resnet50,XceptionNet,MobileNetv2,InceptionResnetV2andVGG16has82.38%,43.75%,81.81%,80.94%,82.35%,89.62%and62.50%accuracyrespectively.Ourmodeloutperformedthenthesemodelswith90.50%accuracy.Thebestresultcorrespondstothemodelweintroduced,basedonthecurrentInceptionNet,VGG19,Resnet50,XceptionNet,MobileNetv2,InceptionResnetV2,andVGG16models.Asaresult,itcanbeconcludedthatthedevelopedframeworkoutperformstheotherones.Theperformanceofallevaluationmodels,asseeninFig.8,indicatethatthemethodissuperiortoothers.Figure9comparestheproposedmodel?saccuracyandlosscurvetothoseofothertraditionalCNNmodels.Fromtheshapeofthecurvesinthismentionedfigure,wecanseethatotherconvolutionalmodelshaveapropensitytooverfitorunderfit,whileourproposedmodelseemstohaveagoodfit.Ourproposedmodel?shighaccuracyandlowlosshavealreadybeenconsideredthebestcasescenarioforanyCNNmodel.Figure10showsthedetectionperformanceofthe 226N.I.Proggaetal.Fig.8.Accuracyofproposedmodel&othertraditionalmodelsproposedmodelandshowsthedetectionresultofthetestdataset?sintwoclasses:venomousandnonvenomous.IncomparisontothepreviouslylistedotherCNNmodels,40imageswerearbitrarilycheckedand9ofthosebeingseeninFig.6.Theidentificationperformanceforthedevelopedmethodisreallyexceedinglydecent.TheconfusionmatricesoftheInceptionNet,VGG19,Resnet50,Xcep-tionNet,MobileNetv2,InceptionResnetV2,andVGG16models,aswellasthesuggestedsystem,areshowninFig.11.ForouradaptedCNNmodel,thediagonalmagnitudeoftheconfusionmatricesinbothclassesisgreaterthantheothermodels.Thatis,theproposedmodelwillpreciselydistinguishthesamenumberoftestsamplesfromourtestdatasetasthecurrentNeuralnetworkmodel.Asaresult,ourmodelsuccessfullyoutperformedtheotherconventionalNeuralnetworkmodelinthisareaaswell.Finally,Fig.12demonstratestheK-foldcrossvalidationresultoftheproposedmodelutilizingthreeoptimizers:Adamax,Adam,andSGD.ByusingadamaxoptimizerforKfoldcrossvali-dationthevalidationaccuracyandthetestingaccuracyforfold-1is83.58%&86.87%,forfold-2itis81.83%&82.52%,forfold-3itis77.44%&80.07%,forfold-4itis84.24%&84.96&andforfold-5itis86.87%&83.20%respectively.Foradamoptimizerthevalidationaccuracyandthetestingaccuracyforfold-1is86.56%&89.22%,forfold-2itis87.57%&88.04%,forfold-3itis83.35%&84.42%,forfold-4itis87.78%&90.48%andforfold-5itis84.36%&88.43%respectively.ForSGDoptimizerthevalidationaccuracyandthetestingaccu-racyforfold-1is84.85%&85.78%,forfold-2itis88.30%&89.67%,forfold-3itis84.78%&85.61%,forfold-4itis83.77%&88.04%,andforfold-5itis87.73%&91.30respectively.TheSGDoptimizerinfold-5producesbetteroutcomesfortheimplementedmodel,aswellastheotherresultsarealsoquitesatisfactory.Alsoitdemonstratesthattheaccuracyoftheclassificationmodelisuna?ectedbythetrainingdata. SnakeClassification227(a)InceptionNet(b)VGG19(c)Resnet50(d)Xception(e)Mobilenetv2(f)Inception-Resnet-v2(g)VGG16(h)ProposedModelFig.9.AccuracyandlosscurveFig.10.Resultofdetectionproducedbytheproposedmodel 228N.I.Proggaetal.(a)InceptionNet(b)VGG19(c)Resnet50(d)Xception(e)Mobilenetv2(f)InceptionResnetv2(g)VGG16(h)ProposedModelFig.11.Confusionmatrixofthemodels(a)adamaxOptimizer(b)AdamOptimizer(c)SGDOptimizerFig.12.K-Foldcrossvalidationforproposedmodelusingdi?erentoptimizer6EpilogueandFutureWorkAnewconvolutionalneuralnetwork-basedarchitecturefordetectingandclassi-fyingvenomousandnon-venomoussnakeswassuggestedduringthiswholestudy.Theframework?sabilitytoacquiresnakefeaturesusingneuralnetworkblocksisclearlydemonstrated.IncomparisontothevariouspossiblythebestConvolu-tionalneuralnetworkframeworksInceptionNet,VGG19,Resnet50,XceptionNet,MobileNetv2,InceptionResnetV2,andVGG16,thearchitecturehasremarkablecategorizationaccuracy.Thisstudylooksathowtodevelopandcreateavenomousandnonvenomoussnakeclassificationmodelthatcouldhelpmankind.Snakevenomouscanbeusedasmedicinaltoolsbydistinguishingbetweenvenomousandnon-venomoussnakes.Snakebitediseasemaybeminimizedbyidentifyingthespeciesofsnakeandadministeringappropriatetreatment. SnakeClassification229Thesuggestedsolutiongreatlyoutshinesstate-of-the-artframeworks,withadramaticincreaseinaccuracyof90.50%,accordingtotheexperimentalresearchreview.Inaddition,themodelperformsadmirablyintermsofKfoldcrossval-idationoutcomes.Itwouldhavebeenmoreusefulinupcominganalysisifweperformagridhuntonthehyper-parametersandfindthemostsuitablenum-berofparametervaluesforKfoldcrossvalidation.Finally,webelievethatthecurrentfindingswouldhopefullyresolveuniquestimulationinrecognizingaddi-tionalsnakeimagesandcategorizingtheminArtificialintelligencebasedenvi-ronments,particularlyinmedicalapplication.Tocreateoursuggestedstructuremorereliableandauthenticated,moreanalysisshouldbedonetoevaluateandgatheradiversedatasetwithsignificantquantitiesofinformationaboutsnakephotographs.Furthermore,manyadditionalNeuralnetworks,suchasDenseNet,EfficientNet,andothers,canbeusedtoenhanceinformationhorizons,andthemodel?soutputcanbecomparedtothatofothermainstreammachinelearningtechniques.Wewilltrytoincludemorepicturesofsnakeinfuturecapturedindi?erentenvironmentandlighteningtoanalyzethemodelperformanceonthoseimages.Alsoanintegrationofdata-driven(CNN)andknowledge-driven(BRBES)approachcanbeproposedtoanalyzeriskassessmentofasnakebiteinthehumanbody[17,19,25,34].References1.Abadi,M.,etal.:Tensorflow:asystemforlarge-scalemachinelearning.In:12th{USENIX}SymposiumonOperatingSystemsDesignandImplementation({OSDI}2016),pp.265?283(2016)2.Abdurrazaq,I.S.,Suyanto,S.,Utama,D.Q.:Image-basedclassificationofsnakespeciesusingconvolutionalneuralnetwork.In:2019InternationalSeminaronResearchofInformationTechnologyandIntelligentSystems(ISRITI),pp.97?102.IEEE(2019)3.Abeysinghe,C.,Welivita,A.,Perera,I.:Snakeimageclassificationusingsiamesenetworks.In:Proceedingsofthe20193rdInternationalConferenceonGraphicsandSignalProcessing,pp.8?12(2019)4.Albawi,S.,Mohammed,T.A.,Al-Zawi,S.:Understandingofaconvolutionalneu-ralnetwork.In:2017InternationalConferenceonEngineeringandTechnology(ICET),pp.1?6.IEEE(2017)5.Annesa,O.D.,Kartiko,C.,Prasetiadi,A.,etal.:Identificationofreptilespeciesusingconvolutionalneuralnetworks(CNN).JurnalRESTI(RekayasaSistemDanTeknologiInformasi)4(5),899?906(2020)6.Bisong,E.:Googlecolaboratory.In:BuildingMachineLearningandDeepLearningModelsonGoogleCloudPlatform,pp.59?64.Springer(2019)7.Bloch,L.,etal.:Combinationofimageandlocationinformationforsnakespeciesidentificationusingobjectdetectionandefficientnets.CLEFworkingnotes(2020)8.Chauhan,R.,Ghanshala,K.K.,Joshi,R.:Convolutionalneuralnetwork(CNN)forimagedetectionandrecognition.In:2018FirstInternationalConferenceonSecureCyberComputingandCommunication(ICSCCC),pp.278?282.IEEE(2018) 230N.I.Proggaetal.9.Chowdhury,R.R.,Hossain,M.S.,UlIslam,R.,Andersson,K.,Hossain,S.:Banglahandwrittencharacterrecognitionusingconvolutionalneuralnetworkwithdataaugmentation.In:2019Joint8thInternationalConferenceonInformatics,Elec-tronics&Vision(ICIEV)and20193rdInternationalConferenceonImaging,Vision&PatternRecognition(icIVPR),pp.318?323.IEEE(2019)10.D?ıaz-Pernas,F.J.,Mart?ınez-Zarzuela,M.,Ant?on-Rodr?ıguez,M.,Gonz?alez-Ortega,D.:Adeeplearningapproachforbraintumorclassificationandsegmentationusingamultiscaleconvolutionalneuralnetwork.In:Healthcare,vol.9,p.153.Multidis-ciplinaryDigitalPublishingInstitute(2021)11.Galvez,R.L.,Bandala,A.A.,Dadios,E.P.,Vicerra,R.R.P.,Maningo,J.M.Z.:Objectdetectionusingconvolutionalneuralnetworks.In:TENCON2018?2018IEEERegion10Conference,pp.2023?2027.IEEE(2018)12.Glorot,X.,Bordes,A.,Bengio,Y.:Deepsparserectifierneuralnetworks.In:Pro-ceedingsofthefourteenthinternationalconferenceonartificialintelligenceandstatistics,pp.315?323.JMLRWorkshopandConferenceProceedings(2011)13.Gulli,A.,Pal,S.:DeepLearningwithKeras.PacktPublishingLtd.,Birmingham(2017)14.Heuvelmans,M.A.,etal.:Lungcancerpredictionbydeeplearningtoidentifybenignlungnodules.LungCancer154,1?4(2021)15.Islam,M.Z.,Hossain,M.S.,UlIslam,R.,Andersson,K.:Statichandgesturerecognitionusingconvolutionalneuralnetworkwithdataaugmentation.In:2019Joint8thInternationalConferenceonInformatics,Electronics&Vision(ICIEV)and20193rdInternationalConferenceonImaging,Vision&PatternRecognition(icIVPR),pp.324?329.IEEE(2019)16.James,A.P.,Mathews,B.,Sugathan,S.,Raveendran,D.K.:Discriminativehis-togramtaxonomyfeaturesforsnakespeciesidentification.HCIS4(1),1?11(2014)17.Jamil,M.N.,Hossain,M.S.,UlIslam,R.,Andersson,K.:Abeliefrulebasedexpertsystemforevaluatingtechnologicalinnovationcapabilityofhigh-techfirmsunderuncertainty.In:2019Joint8thInternationalConferenceonInformatics,Electronics&Vision(ICIEV)and20193rdInternationalConferenceonImaging,Vision&PatternRecognition(icIVPR),pp.330?335.IEEE(2019)18.Jeyaraj,P.R.,Nadar,E.R.S.:Computer-assistedmedicalimageclassificationforearlydiagnosisoforalcanceremployingdeeplearningalgorithm.J.CancerRes.Clin.Oncol.145(4),829?837(2019)19.Kabir,S.,Islam,R.U.,Hossain,M.S.,Andersson,K.:Anintegratedapproachofbeliefrulebaseanddeeplearningtopredictairpollution.Sensors20(7),1956(2020)20.Kamalraj,R.:Deeplearningmodelforidentifyingsnakesbyusingsnakes?bitemarks.In:2020InternationalConferenceonComputerCommunicationandInfor-matics(ICCCI),pp.1?4.IEEE(2020)21.Li,Y.,Hao,Z.,Lei,H.:Surveyofconvolutionalneuralnetwork.J.Comput.Appl.36(9),2508?2515(2016)22.Mahmud,M.,Kaiser,M.S.,McGinnity,T.M.,Hussain,A.:Deeplearninginminingbiologicaldata.Cogn.Comput.13(1),1?33(2021)23.Mahmud,M.,Kaiser,M.S.,Hussain,A.,Vassanelli,S.:Applicationsofdeeplearn-ingandreinforcementlearningtobiologicaldata.IEEETrans.NeuralNetw.Learn.Syst.29(6),2063?2079(2018)24.MohamedAbdEl-Aziz,T.,Soares,A.G.,Stockand,J.D.:Snakevenomsindrugdiscovery:valuabletherapeutictoolsforlifesaving.Toxins11(10),564(2019) SnakeClassification23125.Monrat,A.A.,Islam,R.U.,Hossain,M.S.,Andersson,K.:Abeliefrulebasedfloodriskassessmentexpertsystemusingrealtimesensordatastreaming.In:2018IEEE43rdConferenceonLocalComputerNetworksWorkshops(LCNWorkshops),pp.38?45.IEEE(2018)26.Progga,N.I.,Hossain,M.S.,Andersson,K.:Adeeptransferlearningapproachtodiagnosecovid-19usingx-rayimages.In:2020IEEEInternationalWomeninEngi-neering(WIE)ConferenceonElectricalandComputerEngineering(WIECON-ECE),pp.177?182.IEEE(2020)27.Ren,S.,He,K.,Girshick,R.,Sun,J.:Fasterr-cnn:Towardsreal-timeobjectdetec-tionwithregionproposalnetworks.arXivpreprintarXiv:1506.01497(2015)28.Rezaoana,N.,Hossain,M.S.,Andersson,K.:DetectionandclassificationofskincancerbyusingaparallelCNNmodel.In:2020IEEEInternationalWomeninEngi-neering(WIE)ConferenceonElectricalandComputerEngineering(WIECON-ECE),pp.1?7.IEEE(2020)29.Srivastava,N.,Hinton,G.,Krizhevsky,A.,Sutskever,I.,Salakhutdinov,R.:Dropout:asimplewaytopreventneuralnetworksfromoverfitting.J.Mach.Learn.Res.15(1),1929?1958(2014)30.Terry,P.,Mackway-Jones,K.:Antibioticsinnon-venomoussnakebite.Emerg.Med.J.19(2),142?142(2002).https://emj.bmj.com/content/19/2/142.131.Tosi,S.:MatplotlibforPythonDevelopers.PacktPublishingLtd.,Birmingham(2009)32.T?uske,Z.,Tahir,M.A.,Schl?uter,R.,Ney,H.:Integratinggaussianmixturesintodeepneuralnetworks:softmaxlayerwithhiddenvariables.In:2015IEEEInter-nationalConferenceonAcoustics,SpeechandSignalProcessing(ICASSP),pp.4285?4289.IEEE(2015)33.UddinAhmed,T.,Hossain,M.S.,Alam,M.,Andersson,K.,etal.:AnintegratedCNN-RNNframeworktoassessroadcrack.In:201922ndInternationalConferenceonComputerandInformationTechnology(ICCIT)(2019)34.UlIslam,R.,Andersson,K.,Hossain,M.S.:Awebbasedbeliefrulebasedexpertsystemtopredictflood.In:Proceedingsofthe17thInternationalconferenceoninformationintegrationandweb-basedapplications&services,pp.1?8(2015)35.VanDerWalt,S.,Colbert,S.C.,Varoquaux,G.:Thenumpyarray:astructureforefficientnumericalcomputation.Comput.Sci.Eng.13(2),22?30(2011)36.Warrell,D.A.,etal.:Guidelinesforthemanagementofsnake-bites.Guidelinesforthemanagementofsnake-bites(2010)37.Yanagisawa,H.,Yamashita,T.,Watanabe,H.:AstudyonobjectdetectionmethodfrommangaimagesusingCNN.In:2018InternationalWorkshoponAdvancedImageTechnology(IWAIT),pp.1?4.IEEE(2018)...