introduction, allofthem, havetextualdescriptionsandsomeareassociatedwiththepeopleonthosepictures, automatedmeanstorecognisethefacesontheimagesandconnectthemwithpeoplewould, capabilities, enrich, data, archives, convolutionalneuralnetworkshavebeenthemostsuccessfulapproachtothefacerecognition, probleminthelastdecade, lately, severalpre, trainednetworkshavebeenpublishedandmade, opensource, allthosenetworksweretrainedonalargenumberoffaceimagesgatheredfrom, onlineresourcesliketheinternetmoviedatabaseandgooglesearchresults, thegoalofthe, currentthesisistofindoutwhichofthesenetworkscouldbethebestonetousewiththefaces, inthefotisdatabaseofthenationalarchives, thechallengeisnottrivial, becauseofthe, differenceinthedatathenetworksweretrainedonandthedatawewanttousethemfor, networksweretrainedonmostlyhigherqualitycolorful, redcarpet, photosofcelebrities, photosinthearchivesareoflowerquality, aregrayscaleandcontainmostlyfacesofestonian, politiciansfrompreviousdecades, wherethehaircuts, facialexpressions, andtheageandgender, distributions, inthisthesisthreedifferentneuralnetworksarecompared, theaccuraciesofthenetworks, arecalculatedonacommonbenchmarkdataset, tovalidateoursetupagainsttheknownresults, publishedonthesamedataset, similarbenchmarkdatasetsaregeneratedforthefotis, database, accuracies, networks, calculated, using, methodology, thetheoreticalbackgroundofthisworkisdiscussed, describingartificialneural, networksandfacerecognitioningeneral, convolutionalneuralnetworksaredescribedin, moredetail, thebenchmarkdatasetsaredescribed, thearchitecture, andtrainingmethodologyofthethreenetworksisdescribed, themethodologyand, setupofthebenchmarktestsislaidout, theresultsarepresentedanddiscussed, chapter, contains, conclusions, suggestions, theoretical, definitions, artificialneuralnetwork, isaninformationprocessingmodelinspiredbybiologicalneural, systems, anannismadeupofartificialneurons, thatareusuallyorganizedinlayers, schematic, neural, network, presented, animageofathree, layerneuralnetworkwiththreeinputnodes, nodes, output, image, similarlytoabiologicalneuron, anartificialneurontakesitsinputfromseveralotherneurons, correspondingtosynapticinputinabiologicalneuron, andgeneratesasingleoutput, propagatedbyanaxoninabiologicalneuron, theinputsareweighted, whichmeansthat, differentinputsinfluencetheoutputwithdifferentstrength, agraphicalpresentationofan, artificial, neuron, shown, anartificialneuron, naremultipliedbytheircorresponding, theproductsaresummedandabiasvalueisadded, theresult, ispassedtotheactivationfunction, whichgeneratestheoutputoftheneuron, alayerinanetworkisusuallymadeupofneuronsusingthesameactivationfunction, choiceofanactivationfunctionisveryimportantfortheperformanceoftheneuralnetwork, commontypeofanactivationfunctionissigmoidfunction, afunctionthathasalargeslope, examples, activation, functions, pictured, commonartificialneuronactivationfunctions, noticehowglogisticand, andflattenoffatlargervaluesofx, modified, theneuronsandtheirconnectionsformadirectedgraph, feedforwardneuralnetworkisan, annwherethenodesinthatgraphareconnectedinanon, cyclicalway, bycomparison, recurrentneuralnetworkhascycles, allofthenetworksstudiedinthisthesisarefeedforward, foldcrossvalidationisastatisticalmethodofvalidatingthegeneralisabilityofamodel, howwellthemodelperformswhenpresentedwithnewdata, forthismethodthedatasetthat, isbeingstudiedisdividedintokequal, sizedsets, ineachvalidationrun, oneofthosesetsisleft, outasavalidationfold, themodelistrainedontheremainingk, anditsaccuracytested, onthevalidationfold, thisprocessisrepeatedktimes, usingeachfoldasavalidationfoldonce, results, calculate, average, standard, deviation, alignment, recognition, pipeline, asnotedbyhuangetal, facerecognitioncanbethoughtofaspartofa, detection, eachphaseinthepipelinegeneratestheinputfor, phase, recognitionpipeline, eachphasegeneratesaninput, indetectionphase, thepositionsoffacesinanimagearelocated, usually, thelargestfacefound, isconsideredtobelongtothepersonidentifiedbythatimage, therectangularboundingboxof, thelargestfacefoundisthencroppedfromtheimage, inthealignmentphase, thecropped, imageistransformedtoacanonicalpose, whichgenerallymeansrotatingandtranslatingthe, imagesothattheeyeswouldbelevelandthenoseapproximatelyinthecentreoftheimage, therecognitionphaseinvolvesmatchingthealignedimagetoeitheraknownidentityor, comparing, images, guess, whether, forallimagesinthedatasetsusedinthisstudy, thedetectionpartwasalreadydoneandthe, alignmentalgorithmswereappliedoptionally, themaininterestwascomparingtherecognition, aconvolutionalneuralnetworkisafeedforwardartificialneuralnetworkinspiredbythe, structure, visual, cortex, studied, hubel, wiesel, architectural, features, types, remained, similar, localreceptivefields, eachneuroninalayertakesitsinputfromasmallareaofthe, previouslayer, thisallowsforsmallerfeaturesofanimage, likeedgesandcorners, tobe, recognizedinthefirstlayersofthenetwork, beforecombiningthoseintofeaturesof, higher, abstraction, later, layers, sharedweights, aconvolutionallayerisathree, dimensionalstructureorganizedin, allneuronsinaplanesharethesamesetofweights, becausetheweightsare, shared, allneuronsinaplaneperformthesameoperationontheinputfromtheprevious, layer, andtheoutputoftheneuronsformamappingoffeaturesinthepreviouslayer, whereallofarrowsofthesamecolorrepresentthe, weight, max, poolingoraverage, poolinglayers, neuronsinapoolinglayertaketheirinput, valuesfromsmall, non, overlappingareasfromthepreviouslayerandfind, themaximumoraverageofthevalues, thislessensthedimensionsofthelayerand, provides, invariance, position, anexampleofaconvolutionalfilter, arrowsofthesamecolorrepresent, weights, benchmark, dataset, createdbyhuangetal, labeledfacesinthewild, isadatabaseoffaceimages, commonlyusedforreportingtheperformanceoffacerecognitionalgorithms, theimage, thecreatorsofthedatabasealsodefinedpairsofimagesanddivided, themintosetsmeantfortrainingfacerecognitionalgorithmsandreportingtheirperformance, whileallofthenetworkscomparedinthispaperweretrainedondifferent, largerdatabases, thelfwwasstillusedforperformancereporting, itswideusemakesitwellsuitedfor, techniques, otherversionsofthedatasetarealsoavailable, oneofthemcontainsthesameimagesasthe, originaldataset, buttheimagesaremodifiedusinganautomaticimagealignmenttechnique, examplesofbothoriginalanddeepfunneledimagesarepresentedin, noticehow, funneling, rotated, thedatasetthatwasusedforcomparingtheperformanceofthenetworkisgatheredfromthe, photodatabaseofthenationalarchivesofestonia, human, comparedtolfw, wheretherearemanypeopleofwhomthereisonlyoneimage, theimages, http, www, ra, ee, fotis, areblack, whiteandofvaryingdimensions, examplesofsomeoftheimagesarepresentedin, example, comparison, compared, trainedbytheresearchersofvisualgeometrygroupattheuniversityof, oxford, trainedbymatiisenandtampuuattheinstituteofcomputerscience, universityoftartu, theoriginalversionofthenetworkwastrainedbydongyietal, institute, automation, chinese, academy, sciences, openface, trained, brandon, amos, carnegie, mellon, university, descriptions, theinputisa, architecture, whichtheresearchersgatheredusingthe, process, celebrities, extracted, internet, imdb, collected, google, picturesofallpeoplealreadypresentinthelabeledfacesinthewildandyoutubefaces, removed, twostepsofbothmanualfilteringandautomaticfilteringweredonetoimprovecluster, purity, duplicates, robots, ac, uk, vgg, neuro, cs, ut, https, cmusatyalab, github, io, thecasianetworkhasfivegroupsoftwoconvolutionallayers, thegroupsareseparatedby, thefirstfourpoolinglayersusemaxoperator, thelastpoolinglayerusesaverage, pooling, reluneuronsareusedaftereachconvolutionallayerbutthelast, itdoesnotincludethereluneurons, theinputtothe, originalnetworktrainedbyyietal, usedbothsoftmaxandcontrastiveloss, inthetrainingdone, tartu, softmax, loss, duringthetestsdoneforthisthesis, thedropoutlayerbeingtheonefromwherethefacefeatureswereextracted, similarlytovgg, training, classify, thecasianetworkwastrainedonthecasia, webfacedataset, theopenfacenetworkimplementationbybrandonamosetal, thereisongoingdevelopmentonthe, openfaceimplementation, thestructureofthefacenetarchitectureis, itcontains, amixofregularconvolutionallayers, maxpoolinglayersandinceptionlayers, inceptionlayers, arecomplexconvolutionallayersthatcontainparallelfiltersofdifferentsize, depth, description, available, original, facenet, theopenfacenetworkwastrainedonacombinationoffacescrubandcasia, webface, processing, foreachofthenetworks, adifferentmethodwasusedbythegroupsthattrainedthemfor, aligning, passing, thevggteamusedaprocesscalledoversampling, wherebytheycroppedapartfromeach, cornerandfromthemiddleoftheimagebeingprocessed, theytookthesamecropsfrom, horizontalmirrorversionoftheimage, byone, andaveragedtheresultingfeaturevectors, dalignment, whichgavea, slight, improvement, accuracy, testing, forcasianetwork, theimageswereonlymirroredduringtrainingbeforebeingpassedtothe, foropenface, dlibimageprocessinglibrarywasusedtoaligntheimagessothattheeyesand, approximately, dlib, theexperimentalworkdoneforthisthesisincludedextractingfeaturesforeachthreeofthe, networksandforbothlfwandfotisdatasets, producingthebenchmarkdatasetforthe, analysing, classification, combination, somepreliminaryworkforthelfwdatasetwithbothcasiaandvggnetworkswasalready, theauthorofthispaper, addedsupportforopenfacenetworkandfotisdataset, createdscriptsforpairgenerationand, calculation, evaluated, setup, thetestswererunonalinuxserverwithapythonprogramminglanguageandcaffe, installation, caffeisadeeplearningnetworkimplementedinpython, themodelsofvggand, casiawereimplementedincaffe, openfaceisimplementedintorchdeeplearningtoolkit, whichisbasedonluaprogramminglanguage, fortestingopenface, allofitsrequirements, includingtorchneededtobeinstalled, theexactrequirementsareavailableatopenface, webpage, whileingeneral, goodfacealignmentisconsideredimportantfortherecognitionalgorithms, alignmentalgorithmswereappliedoptionallyduringthetestsinthisstudy, firstly, becausethe, previousresultsandthepreliminarytestsshowedgreatdifferenceinsensitivitytothealignment, betweenthedifferentnetworks, secondly, eachoftheteamsthattrainedthenetworksused, differentalgorithmsforalignment, allofthoseweredifficulttocontrolforandthemaininterest, ofthisthesisliedintherecognitionpartofthedetection, asa, compromise, toprovideanexperimentalcontrolforimagealignment, aseparatesetoftestswith, funneled, themethodologyusedfortestingonthelfwdatasetwasalsoadaptedforfotis, methodologycalculatesaccuracyofpredictingifpairoffacesareofthesamepersonornot, generatingpairsfromthefotisdataset, apythonscriptwaswrittenimplementingthesame, algorithm, generation, outlined, caffe, berkeleyvision, org, toformmatchingpairs, apersonwaschosenatrandomfromallofthepeopleintheset, thereweretwoormoreimagesofthatperson, apairofthosewerechosenatrandom, ifthat, pairwasalreadypresentinthesetofmatchedpairs, theselectionprocesswasstartedagain, added, toformmismatchedpairs, twopeoplewerepickedwithuniformprobability, peoplewith, differentnumberofimageshadthesamechanceofbeingpicked, arandomimageofeachof, thosepeoplewaspicked, iftheresultingpairwasalreadyinthesetofmismatchedpairs, started, thefotisdataset, thenumberwaschosenaftercalculatingthestandarddeviationofthe, cross, validationtestrunswithdifferentnumbersofpairs, andchoosingthenumberofpairsover, whichtherewasnonoticeabledecreaseinthestandarddeviation, theresultofthosetestsare, section, theextractedfacefeaturesinthevgg, andopenfacenetworkarerepresentedby, vectorsofnrealnumbers, thesevectorscanbe, thoughtofaspointsinann, dimensionaleuclideanspace, calculatingtheeuclideandistancebetweenthem, whichiscalculatedasfollows, anb, bn, istance, db, generalization, pythagorean, theorem, calculating, alistoffacepairswasloaded, eithertheonesprovidedwiththelfwdatasetortheones, generated, described, euclidean, distances, theresultingdistancesandknownlabels, whetherthepaircontainedtheimagesofthe, samepersonornot, wererunthroughaten, foldcross, validationprocess, whereineach, validationrun, anoptimalthresholddistancewasfoundonthetrainingfolds, thiswas, theeuclideandistancescalculatedinthelaststepandcalculatingthebinaryclassification, accuracyonthetrainingfolds, thethresholdthatresultedinthemaximalclassification, recorded, theoptimalthresholdsfoundonthetrainingfoldswereusedtocalculatethe, classificationaccuracyoneachcorrespondingvalidationfoldandanaveragewastakenof, withtheoriginal, accuraciesintestsrunbytheauthorandreportedaccuraciesfromthe, creators, reported, tested, whichusedbothsoftmaxand, contrastivelossduringtraining, thenetworkthatwastestedusedonlysofmax, method, suggested, samples, onlyvggnetworksresultcouldbereproducedonthelfwdatasettoaclosemargin, showed, promising, thediscrepancyintheresultsisprobablycausedbyanerrorineitherthe, configurationofthenetworkorthesetupandconfigurationoftherequiredsoftware, nopaperhasbeenpublishedfortheopenface, theinformationforthesetuphadtobegathered, fromsamplescriptsandseparatearticlespublishedontheopenfacewebpage, whichincreases, thechanceofhumanerror, alsonote, thatforlfwandfotisrunsoftheopenfacetests, imageswerealignedbeforebeingpassedtothenetwork, atestwithoutthealignmentwasalso, unpromising, forcasia, thedifferenceintheresultscanprobablybeattributedtothemissingcontrastive, function, andpublished, boththepairgenerationandaccuracycalculationscriptswerecreated, itbecamesimpleto, generatedifferentnumberofpairs, andtesthowthevariationinthecrossvalidationruns, changesifdifferentnumbersofpairsareused, deviations, aftertheseresultstherestofthetestswithfotisdataset, becausenodecreaseinstandarddeviationwasseenwith, thevggnetworkperformedthebestinthetests, bothwhentryingtorepeattheoriginal, resultsaswellasonthefotisdataset, notonlywastheaccuracyonthelfwdatasetthe, greatest, itwasalsoclosesttotheresultspublishedbytheauthorsofthenetwork, whichadds, confidencethattheexperimentwassetupcorrectly, thefotisdatasetispromising, consideringthelowerqualityofthearchiveimages, anditisnot, muchlowerthantheaccuracypublishedbyopenfaceonapresumablyeasierlfwdataset, authorsuggeststoconsiderthevggnetworkasastartingpointforfurtherworkonthefotis, beitclusteringthedataortuningthenetworktogainbetteraccuracy, theopenface, networkshouldnotberuledouteither, theprojectisstillindevelopment, andthelatest, onthecasianetworkcouldalsobepossiblyimprovedmyusingthecontrastivelossmetric, aninterestingobservationisthedifferenceinsensitivitytoimagealignmentthedifferent, networkshave, whilethevggteamreportedonlyamarginaldegradationinperformanceif, imagealignmentwasnotused, inthetestsperformedwithopenfacethedifferencewasfarfrom, marginal, ifthealignmentwasskipped, reminder, onabalanceddatasetofequalnumberofmatchingandmismatchingpairs, arandom, classifier, allthreeofthenetworksusedadifferenttrainingdataset, differentmethodoffacealignment, andhaddifferentnetworkarchitecture, waystotrainanetwork, thatmakesmeta, analysisandcontrollingfordifferentaspectsofthe, experimentsdifficult, theauthorwouldliketoseemorestudieswithstricterexperimental, controlspublished, whereotherindependentvariablesareleftconstantandonlychangestothe, networkarchitecturewouldbemade, thatcouldhelptomovethetheoreticalstudyofartificial, architectures