ANCIENT COINS CERAMICS D AND D DOCUMENTATION FOR PRESERVATION AND RETRIEVAL OF LOST - PDF document

ANCIENTCOINS&CERAMICS-3DAND2DDOCUMENTATIONFORPRESERVATIONANDRETRIEVALOFLOSTHERITAGEHubertMaray,MartinKampelz,FrancoNiccolucciyandRobertSablatnigzyPIN-ServiziDidatticieScienticiperzViennaUniversityofTechnologyl'UniversitdiFirenzeInstituteforComputerAidedAutomationVAST-LaboratoryPatternRecognitionandImageProcessingGroupPiazzaCiardi25,59100Prato,ItalyFavoritenstrasse9/183-2,1040Vienna,Austria hubert.mara@pin.uni.it , niccolucci@uni.it kampel@prip.tuwien.ac.at , sab@prip.tuwien.ac.at http://www.pin.unifi.it http://www.prip.tuwien.ac.at 1INTRODUCTIONDocumentationofceramicsisamaintaskinarchaeology,be-causeceramicsarethemostcommonndings,usedandproducedinlargenumbersbyhumansforseveralthousandsofyears.Ar-chaeologistsuseanalysisofceramics( Leute,1987 2 showingbasicdocumentationfordailyndsandspecializedmethods,whichhelparchaeologiststoanswermorespecicquestionsaboutceramics.Thesecondpart(Section 3 4 ).2CERAMICSThebasisofdocumentationofceramicsisamanuallydrawnhor-izontalintersection,whichiscalledproleline( Leute,1987 ).Astheceramicsarefoundintensofthousandsatvirtuallyev-eryexcavation,thesedrawingsrequirealotoftime,skillandman-powerofexperts.Thereforeweareassistingarchaeologists ininterdisciplinaryprojects( KampelandSablatnig,1999 , Cos-masetal.,2001 )byusinganautomatedsystemforacquisitionanddocumentationofceramicsusinga3D-scannerbasedontheprincipleofstructuredlight( DePieroandTrivedi,1996 , Liska,1999 ).Thissectionisstructuredintothreeparts:Firstwedescribethemethodsusedforbasicdocumentation(Section 2.1 ),whichcanbeappliedoncompleteceramicsandtheirfragments.Thisworkisthebasisforsymmetry-analysisleadingtothedeterminationofthemanufacturingtechniqueofancientceramics,asshowninSection 2.3 .FinallySection 2.4 showsanexampleforprocessingdecoratedceramicsforfurtherarchaeologicalresearch.2.1ACQUISITIONRegardingdocumentationofceramicsthebasicmethodappliedtounbrokenceramicsandfragmentsofceramicsistheestima-tionofaprole-line.Thisisthelongestelongationaround–orcross-sectionthrough–thewallofaceramicdenedbythero-tationalaxis(alsocalledaxisofsymmetry).Thetermrotationalaxisrelatestothefactthatrotationalwheels(plates)havebeenusedforthousandsofyearsformanufacturingceramics.Thisas-sumptioncanbemadeespecially,butnotonlyfordailyndsonarchaeologicalexcavations.Thereforewebaseourworkonusingtherotationalaxistoorientaceramicoritsfragmenttoestimatetheprolelineasitisdonemanuallybydrawings.Furthermorethecreationofsuchmanualdrawingsisatime-consumingtaskrequiringexpert-skillsandoursystemwillhelptodramaticallyreducethetimefordocumentation.Figure 1 1 showsanorientatedfront-viewandamanualdrawingofprolelineofafragmentofaceramic(alsocalledsherd)foundattheexcavationatTelDor,Israel( Stern,1994,extended2000 ). (a) (b)Figure1:(a)Front-viewandmanualdrawing1ofa(b)prole-lineofasherdfoundattheexcavationsatTelDor,Israel. Thereforethechallengingtasksfordevelopingadocumentationsystemforarchaeologyaretobuildasystem,whichisaccurate,portable,inexpensive,easy-to-useandrobustforallkindsoffor-eignclimate,whichcanrangefromdesert,tojungletoarctic.Thismeansseveraltechnologieslikecomputertomographyandotherlaboratoryequipmentisoftennotsuitableforthedailyworkofarchaeologists-especiallynotforceramics. 1ThemanualdrawingwasmadeavailablebyIlanSharon,HebrewUniversityofJerusalem. Asphotographyhasalreadyprovenitsreliabilityforarchaeology,wechosetouseacameraandalight-sourcefor3D-acquisition.Ourrstprototype( Sablatnigetal.,1991 )inourlaboratorywasbasedontheprincipleofstructuredlight( DePieroandTrivedi,1996 ).Itspredecessorwasadigitalstillcamerawithaspe-cialashlightcalledEyetronicsShapeSnatcher( Cosmasetal.,2001 ).Thiscamerawasdevelopedandusedinthe3D-MURALEproject( Cosmasetal.,2001 ).Nowadaysweuse3D-scannersfromtheKonica-MinoltaVivid( Mara,2003 , MaraandHecht,2006 )productrange,becauseoftheirres-olution(0:1mm),whichmeetstherequirementsgivenbyarchaeologistsfortheirdocumentation.Figure 2 showsthesetupofour3D-scannerfromrecentexperimentsattheexcavationsintheValleyofPalpa( ReindelandCuadrado,2001 ),Peru.FurthermoreFigure 2 (a)showsthetriangulationprinciple( Mara,2003 )usingalaser(bottom)andacamera(top)havingawell-knowndistanceandorientation.Additionallytheturntable–alsoshowninthisFigure–isusedtogetacomplete3D-modeloftheceramic.Thenumberof3D-scansdependsonthecomplexityoftheceramicandittypicallyrangesfromtwoscansforsherdsuptoeightscansforvessels.The3D-scansareregisteredusing( Toso-vic,2002 )toreassembeacomplete3D-model. (a)(b)Figure2:Konica-MinoltaVi-9i3D-scanner(a)projectingalaser-plane(bottomarrow)ontoasherd,whilethecamera(top)ac-quirestheprojectionofthelaser.Havingawell-knowndistanceandorientationofthelaser-planeandtheeldofviewofthecam-erathedistance(range)canbeestimated.(b)Detailofthissetupshowingasherdmountedwithplasticineontheturntable,whichisusedforcontrolledacquisitionofallsidesofanobject. 2.2PROCESSINGAftertheregistration,noisefromdustandotherobjectslikehold-ingdevices(e.g.clampsorplasticine)areremovedfromthe3D-model.Thentheorientationisestimatedbasedontheas-sumptionthatceramicsarerotationallysymmetricobjects( Mara,2003 ),becausetheyweregenerallymanufacturedonrotationalplates.Theprincipleofourorientationmethodisttingofcircletemplates( Ganderetal.,1994 ).Incomparisontothecomput-erized,butmanualmethodof( Meleroetal.,2003 )ourorienta-tionmethodcanbeusedfully-andsemi-automatic( Lettneretal.,2006 ).Furthermoreoursystemiscapabletostorethe3D-modelandfurtherarchaeologicalinformation(e.g.description,pho-tographs,etc.)inadatabase.Forsolvingthepuzzlingproblemsofother–typicallyindustriallymanufactured–rotationalobjects,methodslike( PottmannandRandrup.,1998 , Willis,2004 , Or-riols,2004 )canbeapplied.Havinganorientated3D-modelaverticalcross-sectionisestimatedusingthepointofmaximumheightofthe3D-model.thiscross-sectionistheso-calledproleline,whichconcludesthetraditionalarchaeologicaldocumenta-tion.Forcomparisonwiththewellconsolidatedmanualapproachweacquired25sherdsgivenbyarchaeologists.Allthesesherdsare considered”small”,whichmeansthatanysherdsmallerthantheselectedones,cannotbeorientatedmanually.Forfurthercross-validationthesherdshavealsobeenorientatedanddocumentedthroughtheuseofaProlograph( UtiliandDolmazon,2002 ).Forcomparisonoftheprolelinesfromthesethreewaysofdoc-umentation,weusedthenovelmethodofThodberg( Thodberg,2003 )forshapematchingtoestimatethedifferencesbetweenthem.Thishastobedone,becauseforrealdatathereisnogroundtruthingeneral.Figure 3 showsoneestimatedprolelinesusingmanualorien-tation,theProlographandorientationusingcircletemplates.Furtherresultscanareshownin( Mara,2003 ).Concludingtheseresults,wecouldorientate24outofthe25sherds.Comparisonoftheprolelinesbetweenthesystemsshowed,thatthesys-temsagreefor2=3ofthesesherds,whiletheremaining1=3oftheprolelineshaveidenticalshape,butdifferingradiiandorientationinrespecttotherotationalaxis.Thereasonisthatthesesherdsaresmallfragmentsoflargeand/oratvessels(e.g.platescones). (a)(b)Figure3:(a)Prolelinesestimatedbymanualorientation(MN),theProlograph(PR)andusingcircletemplates(3D).Thedottedprolelinesarerotatedand(b)shiftedbytheThod-berg'smethodforcomparison. Assucharathersimpletwo-dimensionalprolelinedoesnotre-ectanyinformationaboutthemanufacturingqualityleadingtothemanufacturingtechnique,wedecidedtoenhanceoursystembygivingthearchaeologistsatooltogatherfurtherinformationabouttheacquired3D-model.Theseenhancementsareshowninthefollowingsections.Figure2.3SYMMETRYANALYSISAsarchaeologistsarealsoexcavatingburialplaces,whereun-brokenceramicsorcompletesetsofsherdsarefound,wearepresentingamethodtodeterminethemanufacturingprocessofceramics,whichrevealsinformationaboutthetechnologicalad-vancementofanancientculture.Furthermorethismethodcanbeapplied,butisnotlimitedto,unbrokenorreconstructedvessels.Technologicaladvancement,isdeterminedbyarchaeologistsbe-tweenceramics,thathavebeenproducedeitheronsloworfastturningrotationalplates.Anotherexampleisanongoingdis-cussionbetweenarchaeologistsabouttheexistenceofrotationalplatesformanufactureofceramicsinSouthAmerica.Thegen-eralopinionisthatinthisregionthewheelwasnotinvented,thereforeceramicswereproducedwithoutarotationalplate(wheel)( WieczorekandTellenbach,2002 )ontheotherhand-sidethereisevidencethatrotationalplateswereused( Carmichael,1986 ).Asweusestructuredlightas3D-acquisitionmethod,wecannotmakeanassumptionsabouttheinternalstructureofaceramicas( WieczorekandTellenbach,2002 ),butwecanestimatethe surfacewithahighresolution(0:1mm).Thereforewecanan-alyzethesymmetryandestimatefeatureslikedeviationofrealsurfacesinrespecttoaperfectlysymmetricalsurface.Suchfea-turescanhelparchaeologiststodecideaboutthetechnologicaladvancementsofancientcultures.Tobeginourinvestigationandanswerquestionsaboutthemanu-facturingprocessofceramics,wechosetousetwomodernpots,whichweremanufacturedinatraditionalway.Thereforethisdatacanbeinterpretedasmixturebetweensyntheticandrealdata,be-causeweusedrealobjects.However,unlikerealarchaeologicalfragments,weknowhowtheywasproduced.Furthermorewedecidedtousethemethodforndingtheorien-tationofasherd( MaraandKampel,2003 ).Webeganwiththeproleline,whichcanbeestimatedinasimilarwaylikeincaseofsherds.Thedifferenceisthatforcompletevesselsthebottomplanecanbeusedfororientation,becauseitisthecounterpartfortherotationalplate,whichdenesthe(orthogonal)axisofrota-tion.Weestimatedmultipleprolelines,whichcanbeoverlaidbytransformingthemintothesamecoordinatesystem,wherethey-axisequalstherotationalaxis.Thereforethedistancebetweenprolelinescanbeestimated.Figure 4 showsthelongestpro-lelineandmultipleprolelinescombinedwiththeside-view,asarchaeologistsshowsuchvesselsintheirdocumentation.Inthecaseofthemultipleprolelines,wehaveestimatedthatthedistancebetweentheprolelinesdiffersandthereforethesepotsandtheirprolelinesareunique.Themaximumdistancebe-tweentwoprolelinesoftherstpotwas9:8mm,whereasforthesecondpotwas21:2mm. (a)(b) (c)(d)Figure4:(a,c)Longestprolelinesand(b,d)multipleprolelinesofmodernceramics,manufacturedintraditionalway,whicharesupposedtobeidentical. InthemultipleprolelinesshownFigure 4 (c,d),thedistancebe-tweenprolelines,measuredparalleltothex-axis,isnotequal.Iftheprolelineswereparallel,thiswouldmean,thatthepotshaveanelliptic(horizontal)cross-section.Asitappears,theasymmetryismorecomplex.Therefore,wechosetoanalyzethe potsslice-by-slicealongtherotationalaxis,presumedasorthog-onaltothebottomplane.Figure 5 (a,c)showshorizontalintersections,whichhavebeenap-pliedwithadistanceof10mmalongtherotationalaxis.Thedistanceof10mmcorrespondstothemanufacturingprocess,whichhaslefttracesintheformofrillsasseenalongtherighthandsidesofFigure 4 (b,d).Theserillsarespaced10mm,whichcorrespondstothewidthofthengerortoolusedto”grow”thepotalongtheaxisoftherotationalplate.Theintersectionsat160mmand170mminheighthavebeendiscarded,astheyin-tersectthe”shoulder”ofthepotwithaverylowangle(5°),resultinginanintersectionhavinganon-representative,randomcurvature.Dividingceramicsintosectionsbycharacteristicpoints(likethe”shoulder”)iscarriedoutbyarchaeologistsforclassication.There-forewechosetoanalyzetheobjectsegmentedintoalowerandanupperpart.Thismeans,wehavetwofragments,whereaxisestimationcanbeapplied,asforsherds(fragments).Theestima-tionoftheaxisisshowninFigure 5 (b,d).Thenumericresultsfortheaxisarethattheyhaveaminimumdistanceof4mmtowardseachotherandtotheaxisdenedbythebottomplane.Further-moretheanglesbetweentheaxesdifferbetween5°to7°. (a)(b) (c)(d)Figure5:(a,c)Top-viewand(b,d)side-viewofthehorizontalcross-sections-thelevelofgraycorrespondstotheheight.Theaxisofrotationforthelowerandupperpartisshownasablackline,denedbythecentersoftheconcentriccircles(shownasdots). Usingtherotationalaxisofthelowerandupperfragment,were-peatedtheestimationoftheprolelines,whichareshowninFig-ure 6 .Themaximumdistancebetweentheprolelineare7mmfortheupperand2mmforthelowerpart.Thereforetherstconclusionisthattheupperandlowerpartsdohaveadifferentaxisofrotation,whichmeansthatthesepartshavebeenproducedseparatelyandcombinedwithouttheuseoftherotationalplate.Wecanconcludethat,basedonthedifferentdeviationofthemul-tipleprolelinesshowninFigure 6 ,theupperpartisoflesserqualitythanthelowerpart.Thisleadstotheconclusionthatthesepartshavebeenmadebypotterswithdifferentexperienceand/oronaslowerrotationalplate.Vice-versathedeviationintheup-perpartofupto7mmcomparedtolessthan2mmofthelowerpart,showsthatafasterturningrotationalplatehasbeenusedthatmoreexperiencewasrequiredformanufacturingtheupperpart. (a)(b) (c)(d) (e)(f) (g)(h)Figure6:Axisofrotationandmultipleprolelinesoftheupperpart(a,e),lowerpart(c,g)and(b,d,f,h)thelongestprolelinesofthepartsoftheobjects. Fromthedifferinganglebetweentheaxisofrotationbasedonthebottomplanecomparedtotheaxisofrotationoftheupperandlowerfragment,wecanconcludethateitherthebottomhasbeenpost-workedorthepotwascontortedbeforebeingredintheoven.Evencorrectingtheaxisforthepartsoftheobject,thehorizontalintersectionsarenotperfectlycircular.Thehorizontalintersec-tionsareelliptic.Thereforeweestimatedthedirectionofthemajorandminoraxisoftheellipses.Weestimatedthattheminoraxishasthesamedirectionastheorientationofthehandle.Thismeansthatthesymmetryofthepotswasbroken,whenthehandlewasattachedandthepotswerestillwet.Figure 7 showthepots,intersectedbyaplane,denedbythecenterofgravityofthepotandthedirectionofthemajoraxisoftheellipses.Theanglebe-tweentheminoraxisandthehandleofthepotwas7°and14°forthesecondpot.Weadditionallyconcludethattheellipsestted( Ganderetal.,1994 )tothehorizontalcross-sectionscanbeusedasanadditionalfeature.Thereforethedistancebetweenthefocioftheellipseisestimated.Ceramicswithadistanceconvergingtowardszero(circularcross-sections)areofhigherquality.Theproposedmethodhasalsobeentestedon17realvessels( Mara (a)(b)Figure7:Planesofsymmetryofthe(a)rstand(b)secondobject. andHecht,2006 ),datedfromtheNASCA-period( Carmichael,1986 ),whichwerefoundintheValleyofPalpa,Peru( ReindelandCuadrado,2001 ).Thereforewecouldseperatetheseves-selsintothreeclassesdeterminedbythesymmetry.Thevessels(60%)oftwoofthesethreeclasseswerenotproducedonrota-tionalplates.Besidethisinformation–answertothequestionofthemanufacturingtechnique–abouttheuseofrotationalplatesinSouthAmerica,thisclassicationisusedbyarchaeologistsoftheGermanInstituteforArchaeology(DAI,Bonn)forrenementoftheirclassicationschemes.2.4LINEDETECTIONONSURFACESTheapproachshowninthissectionisanexampletoattemptingtoaddtheanalysisofcolorinformationofceramics,whichalreadyhavebeenacquiredforautomatedproleestimationasshownintheprevioussections.Theonlyrequirementistoutilizea3D-scannercapableoftexture-mapping.Thegoalofthisexampleistoapplyimageprocessingmethodsdirectlytothistypeofsurface(texturemapped)imagesinordertoanalyzethedrawingspaintedonsurfacesofthepottery.Thereareseveralapproachesintheliteraturethatdealwithfea-turedetectiononsurfaces.( R¨ossletal.,2000 )appliesmorpho-logicaloperatorstoextractfeaturelinesontriangulatedsurfaces.Asimilarapproachwaspresentedby( Pottmannetal.,2004 )whoapplymorphologicaloperatorsonsurfaces,usingnewmetric,calledisophoticmetric.Whilethemaingoaloftheseapproachesistodiscovercharacteristicsinthecurvatureofthesurface(e.g.defectsonsurfaces),weareinterestedinthetexture,i.e.thepic-torialinformationrepresentedbythetexture.Amethodtoenhancelinearstructuresin2D-imagesisthecon-volutionofanimagewithanisotropicsecondderivativelter(LoG)( MarrandHildreth,1979 ).Anisotropiclterhasthead-vantage,thatnodirectionalinformationfromtheinputdataisnecessary.Ontheotherhand,nodirectionalinformationispro-videdbytheoutput.Inourcase,theadaptationfor3D-dataisstraightforwardandisaccomplishedintwosteps:Step1:Denitionofthe”convolutionwindow”Theimageonthesurfaceisgivenbyameshofverticesviandeachvertexisassignedagreyvaluegi.Foreveryvertexviasetofneighboredverticesbi,i.e.theconvolutionwindow,isestimated.Thisisachievedbyestimationofthedistancesalongthesurface(calledgeodesicdistances)todirectneighborsofvi.Thesetofver-ticesofthedirectneighborsofviiscalled1-ring.For1-ringsthegeometricdistanceequalsthegeodesicdistances.So,basedonthegeodesicdistancesofthe1-ring,thegeodesicdistancestoviforallneighborsofthe1-ringcanbeapproximated( Sun andAbidi,2001 ).Thisapproximationisiterativelycontinuedas”marchingfront”( NovotniandKlein,2002 )untilnomoreverticeswithageodesicdistancelowerthandistarefound.AschemaofthegeodesicdistanceisdepictedinFigure 8 (a).Fig-ure 8 (c)showsadetailedversionofonerandomlychosenvertexvianditsgeodesicneighborhood.ThegreylevelinFigure 8 cor-respondstothegeometricdistances,darkermeanslowergeodesicdistance.Step2:Theconvolution:Thedistanceofavertexvitobiisdenotedasdij.The1D-lterkernelisdenedas:f(x)=1 2x2 2�1e�x2 22Thediscreteconvolutionisimplementedasthesumofacomponent-wisemultiplicationofthegreyvaluesgjassignedtoallneigh-boredverticesbiandviwiththecorrespondingvalueoff(di),ci=Pj2bi;vigjf(dij).Theacquisitionsystemprovidestwotypesofdata,3D-rangedataand2D-imagesofthesurface.Bothdatasetscanberelatedtoeachother,e.g.eachvertexofthe3Dmeshcanbeassignedacolorvalueandvice-versa.Forourexperimentsweusedtheintensityvalue,i.e.theIcomponentintheHSIcolorspace,calculatedfromthecolortexturevalue.FortheapplicationofthederivativeGaussianlterweselectallneighboringverticeswithageodesicdistancelessthan1:5thewidthoftheexpectedline.Thisisanempiricalvalue,whichwasobtainedfromourexperimentswith2D-lineimages,wherethesizeofthelterker-nelhadtobe3timesthewidthofthelines.Figure 8 showsthecomparisonbetween(c)themanualdrawing,(d)theautomatedunwrapping( doCarmo,1976 , Strubecker,1969 )ofthetexture-mapofthe3D-modeland(e)thedetectedlineshighlightedinblack.Regardingperformancethisexperimentshowsthattheau-tomatedestimationofadrawingsimilartothemanualdrawingcanbeobtainedwithinsecondsinsteadofhours.Futureworkwilladdpatternrecognitionmethods(e.g.shape-matching)toobtainmissingpartsofthepaintingsandtoautomaticallyndsimilardecorationswithinacollectionof3D-models.3COINSAswedonotwanttolimitoursystemtoceramics,werecentlystartedresearchforintegratingthedocumentationofancientcoins,anothercommonnd,whichistypicallyacquiredbydigitalcam-eras.Besidesautomatedcoinrecognitionandclassicationbasedon2D-images,wealsowanttoacquire3D-data,becauseme-dievalcoinsarenotasatastheirmoderncounterparts.Thiswouldleadtonewperspectivesofstudyingancientcoinsaswellasnewstrategiesforrepresentingthem.Basedontheimagesofcoinsfordocumentationpurposes,weproposeapatternrecogni-tionsystemforautomaticestimationofanabstractdescriptionandfordeterminationofthetypeanancientcoinbelongsto.Whilethiscanbedonequitequicklybyanexpert,wewereaskedbythemtoapplythissystemonimagesfoundonthewebusingweb-crawlingtechniques,becausecoinsareoftenstolenfromex-cavationsandsoldontheblackmarketusingtheinternet.Recentresearchapproachesforcoinclassicationalgorithmsfo-cussolelyontherecognitionofmoderncoins.Appliedpatternrecognitionalgorithmsaremanifoldrangingfromneuralnetworks( M.Fukumietal.,1992 , Bremananthetal.,2005 )toeigenspaces( Huberetal.,2005 ),decisiontrees( Davidsson,1996 ),edgede-tectionandgradientdirections( N¨olleetal.,2003 , Reisertetal., (a) (b) (c) (d) (e)Figure8:(a)Graphicillustratingtheactivevertexv1(black)andtheneighboredverticesb1withinthegeodesicdistancedist.(b)Detailedsampleofarandomlyselectedvertexvanditsgeodesicneighborhood.Thelevelofgreycorrespondstothegeodesicdis-tance-darkermeanscloser.Unwrappingofthedecorationofvessel2801-V3oftheexcavationintheValleyofPalpa,Peru:(c)manualdrawing,(d)unwrappedtextureand(e)texturewithdetectededges(highlightedinblack). 2006 ),andcontourandtexturefeatures( vanderMaatenandPostma,2006 ).( Huberetal.,2005 )presentsamultistageclas-sierbasedoneigenspacesthatisabletodiscriminatebetweenhundredsofcoinclasses.Duetothecontrolledsetupofthesys-tempresentedcoindetectionbecomesatrivialtask.Theyreportcorrectclassicationfor92.23%ofall11;949coinsinasam-pleset.( vanderMaatenandPoon,2006 )presentsacoinclas-sicationsystembasedonedge-basedstatisticalfeatures,calledCOIN-O-MATIC.Thesystemissubdividedintovestages:seg-mentation,featureextraction,preselection,3-nearestneighborapproach,andverication.AttheMUSCLECISCoinCompeti-tion( N¨olleetal.,2006 )themethodachievedarecognitionrateof67.31%onabenchmarksetof10;000coins.Thecoinclassica-tionmethodproposedby( Reisertetal.,2006 )isbasedongradi-entinformation.Similartotheworkof( N¨olleetal.,2003 )coinsareclassiedbyregisteringandcomparingthecoinwithapres-electedsubsetofallreferencecoins.TheproposedmethodwontheMUSCLECISCoinCompetition2006( N¨olleetal.,2006 )witharecognitionrateof97.24%onabenchmarksetof10;000coins.Testsperformedonimagecollectionsbothofmedievalandmod-erncoinsshowthatalgorithmsperformingwellonmoderncoinsdonotnecessarilymeettherequirementsforclassicationofme-dievalones( vanderMaatenandPostma,2006 ).Themaindiffer-encebetweenancientandmoderncoinsisthattheancientcoinshavenorotationalsymmetryandconsequentlytheirdiameterisunknown.Sinceancientcoinsarealltooofteninverypoorcon-ditions,commonrecognitionalgorithmscaneasilyfail.Thefea-turesthatmostinuencethequalityoftherecognitionprocessareyetunexplored.Anexampleofanancientcoinfromadatasetconsistingof3;000highresolutionimagesofancientcoinsonconstant,whitebackgroundisshowninFigure 9 2 .Thecoinspic-tureRomanemperorsfrom30B.C.to300A.D.whoformthe106classesofthedataset.Intherststageofourrecentlyinitiatedresearch,westartedbyanalyzingandevaluatingtheexistingalgorithmsforcoinrecogni- 2ThedatasetwasmadeavailablebyKlausVondrovec,NumismaticCommission,Vienna,Austria Figure9:Exampleimagesoftheancientdataset.1 tion.InitaltestswereperformedusingaMUSCLECISdatabaseofcoinsconsistingofover9;100trainingimages(distributedover100classesofcoins)and1;100testimages.WeappliedthealgorithmprovidedbyMaaten( vanderMaatenandPostma,2006 )basedonedge-basedstatisticaldistributions.UnlikeMaaten,wecompletelyignoredsensormeasures(suchasthicknessordi-ameter)andthusweonlyrelyontheextractedfeatures.Ourex-perimentsbasedonthemulti-scaleedgeangle-distancedistribu-tionachievedarecognitionrateof63%.Surprisingly,coinclas-sicationbasedsolelyonareameasurementachievedarateof45%.Ancientcoinsdifferstronglyfrommodernones.Crucialinuencehasthenatureofthemedievalcoins,whicharelessde-tails,norotationalsymmetryandpoorconditionsduetoabrasionorfouling.Hence,coinclassicationbasedsolelyonedge-basedfeatureswouldmostprobablyfailwiththeclassicationofan-cientcoins.Inthenextstage,wewillperformtestsonamedievaldatabasetondthosefeatures(orsetoffeatures)thatmostinu-encethequalityofmedievalcoinrepresentations.4SUMMARYANDOUTLOOKSummarizingthepresentworkshowshowourheritagecanberetrieved,whenithasbeenlostorstolenandhowitcanbepre-servedusingdigitaldocumentationusing3D-and2D-imagepro-cessingandpatternrecognitionmethods.Furthermoretheseau-tomatedmethodshelpexpertsofculturalheritagetodramaticallyreducetimefordocumentationofthemajorityofndswithinar-chaeologyandtoreducethetimetosearchnumismaticdatabasescontainingthousandsofcoins.Besideimprovementofexistingmethodsfor3D-Visionanotherimportantworkhasrecentlybeguntoensuretheintellectualin-tegrity,reliability,transparency,documentation,standards,sus-tainabilityandaccessibilityoftheinformationgatheredbytheincreasinguseof3D-scanners.ThereforeweareadoptingTheLondonCharter( Beachametal.,2006 ),whichwillbeafuturestandardfortheuseof3D-VisionwithinCulturalHeritage.Forfutureworkwewillcombinethemethodologiesforcoinsandceramicstoachieveamoregenericsystemtoretrievestolence-ramicsandtodocumentandmanagecoinswithindigitallibraries.Furthermorethemethodsusedtoretrievestolenndswillalsobeusedtorevealnewknowledgehiddenwithinthetremendousamountsofndsusingfeaturesoftheirsurfaces.ACKNOWLEDGEMENTSTheauthorswouldliketothank PaulKammerer ,ViennaUni-versityofTechnology, InstituteforAutomation,PRIPGroup forassistancewiththeadoptionofthemethodsofline-detectionto3D-models.Furthermorewewouldliketothankallthearchaeologicalinsti-tutesrecentlyinvolvedinourworkforprovidingobjects,newchallengesandfeedback.Thesearchaeologicalinstitutesare: The(German)ArchaeologicalInstitute( DAI,Bonn ),theAustrianGovernmentAgencyforCulturalHeritage( Bundesdenkmalamt ),theMuseumforHistoryofArts( KunsthistorischesMuseum )inViennaandtheAustrianAcademyofSciences( ¨OAW ).FinallywewouldliketothanktheEuropeanCommissionforthesupportofthepresentedworkdonewithinthe CHIRON -fellowship(MEST-CT-2004-514539)andtheCOINS-project(FP6-SSP5-044450).REFERENCES Beacham,R.,Denard,H.andNiccolucci,F.,2006.Anin-troductiontothelondoncharter.In:M.I.etal.(ed.),Thee-volutionofInformationCommunicationTechnologyinCul-turalHeritage:wherehi-techtouchesthepast:risksandchal-lengesforthe21stcentury,ShortpapersfromthejointeventCIPA/VAST/EG/EuroMed,Archaeolingua,Budapest,Hungary. Bremananth,R.,Balaji,B.,Sankari,M.andChitra,A.,2005.Anewapproachtocoinrecognitionusingneuralpatternanalysis.In:Proc.ofIEEEIndicon2005Conference,pp.366–370. Carmichael,P.,1986.NASCAPotteryConstruction.In:NawpaPacha,Vol.24,Berkeley,California. Cosmas,J.,Itagaki,T.,Green,D.,Grabczewski,E.,Gool,L.V.,Zalesny,A.,Vanrintel,D.,Leberl,F.,Grabner,M.,Schindler,K.,Karner,K.,Gervautz,M.,Hynst,S.,Waelkens,M.,Pollefeys,M.,DeGeest,R.,Sablatnig,R.andKampel,M.,2001.3DMU-RALE:AMultimediaSystemforArchaeology.In:ProceedingsoftheInternationalConferenceonVirtualReality,ArchaeologyandCulturalHeritage,Athens,Greece,pp.297–305. Davidsson,P.,1996.Coinclassicationusinganoveltechniqueforlearningcharacteristicdecisiontreesbycontrollingthedegreeofgeneralization.In:Proc.of9thInt.ConferenceonIndustrial&EngineeringApplicationsofArticialIntelligence&ExpertSystems(IEA/AIE-96),pp.403–412. DePiero,F.W.andTrivedi,M.M.,1996.3-DComputerVisionUsingStructuredlight:Design,Calibration,andImplementationIssues.AdvancesinComputers43,pp.243–278. doCarmo,M.P.,1976.DifferentialGeometryofCurvesandSurfaces.Prentice-Hall. Gander,W.,Golub,G.andStrebel,R.,1994.Least-squaresttingofcirclesandellipses.BIT34,pp.558–578. Huber,R.,Ramoser,H.,Mayer,K.,Penz,H.andRubik,M.,2005.Classicationofcoinsusinganeigenspaceapproach.Pat-ternRecognitionLetters26(1),pp.61–75. Kampel,M.andSablatnig,R.,1999.On3dModellingofAr-chaeologicalSherds.In:ProceedingsoftheInternationalWork-shoponSynthetic-NaturalHybridCodingandThreeDimen-sionalImaging,pp.95–98. Lettner,M.,Mara,H.,M¨uller,A.,Sablatnig,R.,Singer,M.andKrenn,M.,2006.PAT:ProleAnalysisToolfortheDocumen-tationofArchaeologicalFinds.In:Proc.ofElectronicImaging&theVisualArt(EVA'06Vienna)-DigitalCulturalHeritage-EssentialforTourism,pp.83–90. Leute,U.,1987.Archaeometry:AnIntroductiontoPhysicalMethodsinArchaeologyandtheHistoryofArt.JohnWiley&Sons. Liska,C.,1999.DasAdaptiveLichtschnittverfahrenzurOber¨achenkonstruktionmittelsLaserlicht.Master'sthesis,Vi-ennaUniversityofTechnology,ViennaUniversityofTechnol-ogy,InstituteofComputerAidedAutomation,PatternRecogni-tionandImageProcessingGroup. Mara,H.,2003.Documentationofrotationallysymmetricalar-chaeologicalndsby3dshapeestimation.TechnicalReportPRIP-TR-103,ViennaUniversityofTechnology,Inst.ofCom-puterAidedAutomation,PatternRecognitionandImageProcess-ingGroup. Mara,H.andHecht,N.,2006.3D-AcquisitionandAnalysisoffreehandmanufacturedNASCACeramics.In:Proc.ofCAA06:ComputerAplicationsanQuantitativeMethodsinArchaeology,p.toappear. Mara,H.andKampel,M.,2003.AutomatedExtractionofPro-lesfrom3DModelsofArchaeologicalFragments.In:O.Altan(ed.),Proc.ofCIPA2003:XIXCIPAInt.Symposium:NewPer-spectivestoSaveCulturalHeritage,CIPA2003organisingcom-mitte,pp.87–93. Marr,D.andHildreth,E.,1979.Theoryofedgedetection.In:Proc.R.Soc.Lond.,B204,pp.301–328. Melero,F.,Leon,A.,Contreras,F.andTorres,J.,2003.Anewsystemforinteractivevesselreconstructionanddrawing.In:Pro-ceedingsofCAA'03:ComputerApplicationsinArchaeology,pp.8–12. M.Fukumi,S.Omatu,F.TakedaandT.Kosaka,1992.Rotation-invariantneuralpatternrecognitionsystemwithapplicationtocoinrecognition.IEEETransactionsonNeuralNetworks3(2),pp.272–279. N¨olle,M.,Penz,H.,Rubik,M.,Mayer,K.J.,Holl¨ander,I.andGranec,R.,2003.Dagobert–anewcoinrecognitionandsortingsystem.In:Proc.ofthe7thInternationalConferenceonDigitalImageComputing-TechniquesandApplications(DICTA'03),MacquarieUniversity,Sydney,Australia.,CSIROPublishing,pp.329–338. N¨olle,M.,Rubik,M.andHanbury,A.,2006.Resultsofthemus-cleciscoincompetition2006.In:Proc.oftheMuscleCISCoinCompetitionWorkshop,Berlin,Germany,pp.1–5. Novotni,M.andKlein,R.,2002.Computinggeodesicdistancesontriangularmeshes.In:Inproc.ofthe10thIntl.ConferenceinCentralEuropeonComputerGraphics,VisualizationandCom-puterVision2002(WSCG02). Orriols,X.,2004.GenerativeModelsforVideoAnalysisand3DRangeDataApplications.Ph.D.thesis.UniversitatAutonomadeBarcelona,Spain. Pottmann,H.andRandrup.,T.,1998.Rotationalandhelicalsurfaceapproximationforreverseengineering.Computing60,pp.307–322. Pottmann,H.,Steiner,T.,Hofer,M.,Haider,C.andHanbury,A.,2004.Theisophoticmetricanditsapplicationstofeaturesen-sitivemorphologyonsurfaces.In:ComputerVision-ECCV2004,PartIV,volume3024ofLectureNotesinComputerSci-ence,Springer,pp.560–572. Reindel,M.andCuadrado,J.,2001.LosMolinosundLaMuna.ZweiSiedlungszentrenderNasca-KulturinPalpa,S¨udperu.In:P.vonZabern(ed.),InBeitr¨agezurAllgemeinenundVergle-ichendenArch¨aologie,Vol.21. Reisert,M.,Ronneberger,O.andBurkhardt,H.,2006.Anef-cientgradientbasedregistrationtechniqueforcoinrecognition.In:Proc.oftheMuscleCISCoinCompetitionWorkshop,Berlin,Germany,pp.19–31. R¨ossl,C.,Kobbelt,L.andSeidel,H.,2000.Extractionoffea-turelinesontriangulatedsurfacesusingmorphologicaloperators.In:Proc.SmartGraphics2000,AAAISpringSymposium,AAAIPress. Sablatnig,R.,Menard,C.andDintsis,P.,1991.APreliminaryStudyonMethodsforaPictorialAcquisitionofArchaeologi-calFinds.TechnicalReportPRIP-TR-010,ViennaUniversityofTechnology,InstituteofComputerAidedAutomation,PatternRecognitionandImageProcessingGroup. Stern,E.,1994,extended2000.Dor-RuleroftheSeas.IsraelExplorationSociety. Strubecker,K.,1969.Differentialgeometrie:TheoriederFl¨achenmetrik.SammlungG¨oschen,Vol.1197,DeGruyter,Berlin. Sun,Y.andAbidi,M.,2001.SurfaceMatchingby3DPoint'sFingerprint.In:Proc..EighthIEEEInternationalConferenceonComputerVision(ICCV'01),Vol.2,pp.263–269. Thodberg,H.,2003.MinimumDescriptionLengthShapeandAppearanceModels.Vol.2732. Tosovic,S.,2002.Adaptive3DModelingofObjectsbycom-biningShapefromSilhouetteandShapefromStructuredLight.Master'sthesis,ViennaUniversityofTechnology,ViennaUni-versityofTechnology,InstituteofComputerAidedAutomation,PatternRecognitionandImageProcessingGroup. Utili,F.andDolmazon,P.,2002.BemerkungenzudreiattischenTrinkgef¨aßenimReiss-MuseuminMannheim.In:M.Bentz(ed.),VasenforschungundCorpusvasorumantiquorum:Stan-dortbestimmungundPerspektiven,BeiheftezumCVA1,p.37. vanderMaaten,L.J.andPoon,P.,2006.Coin-o-matic:Afastsystemforreliablecoinclassication.In:Proc.oftheMuscleCISCoinCompetitionWorkshop,Berlin,Germany,pp.07–18. vanderMaaten,L.J.andPostma,E.O.,2006.Towardsautomaticcoinclassication.In:Proc.oftheEVA-Vienna2006,Vienna,Austria,pp.19–26. Wieczorek,A.andTellenbach,M.,2002.ExkurszurFragederDrehscheibenkeramik.In:AndieM¨achtederNatur-MythenderaltperuanischenNasca-Indianer.KatalogzurAusstellungimReiss-Engelhorn-Museum.Mainz:PhilippvonZabern.,pp.54–63. Willis,R.,2004.Stochastic3DGeometricModelsforClassica-tion,Deformation,andEstimation.Ph.D.thesis.BrownUniver-sity,RhodeIsland,USA.