EarBiometrics: ASurveyofDetection,FeatureExtractionandRecognitionMethodsAnikaPug,ChristophBusch July2,2012The possibility of identifying people by the shape of their outer ear was rstdiscovered by the French criminologist Bertillon, and rened by the AmericanpoliceocerIannarelli, whoproposedarstearrecognitionsystembasedononlysevenfeatures.Thedetailedstructureoftheearisnotonlyunique,butalsopermanent,astheappearanceof theeardoesnotchangeoverthecourseof ahumanlife. Additionally, theacquisitionofearimagesdoesnotnecessarilyrequireapersonscooperationbutisneverthelessconsideredtobenon-intrusivebymostpeople.Because of these qualities, the interest in ear recognition systems has grownsignicantly in recent years. In this survey, we categorize and summarize ap-proachestoeardetectionandrecognitionin2Dand3Dimages. Then, weprovideanoutlookoverpossiblefutureresearchintheeldof earrecogni-tion,inthecontextofsmartsurveillanceandforensicimageanalysis,whichweconsidertobethemostimportantapplicationofearrecognitioncharac-teristicinthenearfuture.1 IntroductionAsthereisanever-growingneedtoautomaticallyauthenticateindividuals, biometricshas been an active eld of research over the course of the last decade. Traditional meansof automatic recognition,such as passwords or ID cards,can be stolen,faked,or forgot-ten. Biometric characteristics, on the other hand, are universal, unique, permanent, andmeasurable.The characteristic appearance of the human outer ear (or pinna) is formed by the outerhelix,theantihelix,thelobe,thetragus,theantitragus,andtheconcha(seeFigure1).HochschuleDarmstadt-CASED,Haardtring100,64295Darmstadt,Germany1This paper is a preprint of a paper accepted by IET Biometrics and is subject to Institution of Engineering and Technology Copyright. When the final version is published, the copy of record will be available at IET Digital LibraryThenumerousridgesandvalleysontheouterearssurfaceserveasacousticresonators.For lowfrequencies thepinnareects theacousticsignal towards theear canal. Forhighfrequenciesitreectsthesoundwavesandcausesneighbouringfrequenciestobedropped. Furthermoretheouterearenableshumanstoperceivetheoriginodasound.The shape of the outer ear evolves during the embryonic state from six growth nodules.Its structure, therefore, is not completely random, but still subject to cell segmentation.Theinuenceofrandomfactorsontheearsappearancecanbestbeobservedbycom-paringtheleftandtherightearofthesameperson. Eventhoughtheleftandtherightearshowsomesimilarities,theyarenotsymmetric[1].Figure1:Characteristics of the human ear the German criminal police uses for personalidenticationofsuspectsTheshapeoftheouterearhaslongbeenrecognizedasavaluablemeansforpersonalidentication by criminal investigators. The French criminologist Alphonse Bertillon wasthe rst to become aware of the potential use for human identication through ears, morethan a century ago [2]. In his studies regarding personal recognition using the outer ear in1906,RichardImhoferneededonlyfourdierentcharacteristicstodistinguishbetween500dierent ears [3]. Startingin1949, theAmericanpoliceocer AlfredIannarelliconductedtherstlargescalestudyonthediscriminativepotential of theouterear.Hecollectedmorethan10000earimagesanddetermined12characteristicsneededtounambiguouslyidentifyaperson[4]. Iannarelli alsoconductedstudies ontwins andtriplets, discovering that ears are even unique among genetically identical persons. EventhoughIannarellisworklacksacomplextheoreticalbasis,itiscommonlybelievedthatthe shape of the outer ear is unique. The studies in[5] and[6] showthat all earsof theinvestigateddatabases posses individual characteristics, whichcanbeusedfordistinguishingbetweenthem. Becauseof thelackof asucientlylargeeardatabase,thesestudiescanonlybeseenashints,notevidence,fortheouterearsuniqueness.Research about the time-related changes in the appearance of the outer ear has shown,thattheearchangesslightlyinsizewhenapersonages[7] [8]. Thisisexplainedbythefactthatwithageingthemicroscopicstructureoftheearcartilagechanges, which2reducestheskinelasticity. Arststudyontheeectof shortperiodsof timeonearrecognition [9] shows that the recognition rate is not aected by ageing. It must, however,bementionedthatthelargesttimeelapsingdierenceinthisexperimentwasonly10months, anditthereforeisstill subjecttofurtherresearchwhethertimehasacriticaleectonbiometricearrecognitionsystemsornot.The ear can easily be captured from a distance, even if the subject is not fully cooper-ative. Thismakesearrecognitionespeciallyinterestingforsmartsurveillancetasksandforforensicimageanalysis. Nowadaystheobservationof characteristicsisastandardtechniqueinforensicinvestigationandhasbeenusedasevidenceinhundredsofcases.Thestrengthofthisevidencehas, however, alsobeencalledintoquestionbycourtsinthe Netherlands [10]. In order to study the strength of ear prints as evidence, the Foren-sicEaridenticationProject(FearID)wasinitiatedbynineinstitutesfromItaly, theUK,andtheNetherlandsin2006. Intheirtestsystem, theymeasuredanEqualErrorRate (EER) of 4% and came to the conclusion that ear prints can be used as evidence inasemi-automatedsystem[11]. TheGermancriminalpoliceusethephysicalpropertiesoftheearinconnectionwithotherappearance-basedpropertiestocollectevidencefortheidentityofsuspectsfromsurveillancecameraimages. Figure1illustratesthemostimportant elements and landmarks of the outer ear, which are used by the German BKAformanualidenticationofsuspects.Inthisworkweextendexistingsurveysonearbiometrics,suchas[12],[13],[14],[15]or[16]. Abazaet al. [17] contributedanexcellentsurveyonearrecognitioninMarch2010. Theirworkcoversthehistoryofearbiometrics,aselectionofavailabledatabasesandareviewof 2Dand3Dearrecognitionsystems. ThisworkamendsthesurveybyAbazaetal. withthefollowing: Asurveyoffreeandpubliclyavailabledatabases. More than 30 publications on ear detection and recognition from 2010 to 2012 thatwerenotdiscussedinoneoftheprevioussurveys. Anoutlookoverfuturechallengesforearrecognitionsystemswithrespecttocon-creteapplications.In the upcoming Section we give an overview of image databases suitable for studyingeardetectionandrecognitionapproachesfor2Dand3Dimages. Thereafter,wediscussexistingeardetectionapproacheson2Dand3Dimages. InSection4wegoontogiveanoverviewof earrecognitionapproachesfor2Dimages, andinSection5wedothesamefor3Dimages. Wewill concludeourworkbyprovidinganoutlookoverfuturechallengesandapplicationsforearrecognitionsystems.2 AvailableDatabasesforEarDetectionandRecognitionInordertotestandcomparethedetectionorrecognitionperformanceof acomputervisionsystem, ingeneral, andabiometric systeminparticular, image databases of3sucientsizemustbepubliclyavailable. Inthissection, wewanttogiveanoverviewof suitabledatabasesforevaluatingtheperformanceof eardetectionandrecognitionsystems, which can either be downloaded freely or can be licensed with reasonable eort.2.1 USTBDatabasesThe University of Science and technology in Beijing oers four collections (http://www1.ustb.edu.cn/resb/en/doc/Imagedb\_123\_intro\_en.pdf) (http://www1.ustb.edu.cn/resb/en/doc/Imagedb\_4\_intro\_en.pdf) of 2D ear and face prole images to theresearchcommunity. AllUSTBdatabasesareavailableunderlicense. DatabaseI: Thedatasetcontains180imagesintotal, whichweretakenfrom60subjectsin3sessionsbetweenJulyandAugust2002. Thedatabaseonlycontainsimagesof therightearfromeachsubject. Duringeachsession, theimagesweretaken under dierent lighting conditions and with a dierent rotation. The subjectswerestudentsandteachersfromUSTB. Database II: Similarly to database I, this collection contains right ear images fromstudentsandteachersfromUSTB. Thistime, thenumberof subjectsis77andtherewere4dierentsessionsbetweenNovember2003andJanuary2004. Hencethe database contains 308 images in total, which were taken under dierent lightingconditions. Database III: Inthis dataset 79, students andteachers fromUSTBwere pho-tographed in dierent poses between November 2004 and December 2004. Some ofthe ears are occluded by hair. Each subject rotated his or her head from 0 degreesto60degreestotherightandfrom0degreesto45degreestotheleft. Thiswasrepeatedontwodierentdaysforeachsubject, whichresultedin1600imagesintotal. DatabaseIV: Consistingof 25500imagesfrom500subjectstakenbetweenJune2007andDecember 2008, this is the largest dataset at USTB. The capturingsystemconsistsof17camerasand,iscapableoftaking17picturesofthesubjectsimultaneously. Thesecamerasaredistributedinacirclearoundthesubject,whois placedinthecenter. Theinterval betweenthecameras is 15degrees. Eachvolunteerwasaskedtolookupwards, downwardsandeyelevel, whichmeansthatthisdatabasecontainsimagesatdierentyawandpitchposes. Pleasenotethatthisdatabaseonlycontainsonesessionforeachsubject.2.2 UNDDatabasesThe University of Notre Dame (UND) oers a large variety of dierent image databases,which can be used for biometric performance evaluation. Among them are ve databasescontaining 2D images and depth images, which are suitable for evaluation ear recognitionsystems. All databasesfromUNDcanbemadeavailableunderlicense(http://cse.nd.edu/~cvrl/CVRL/Data\_Sets.html).4 CollectionE:464rightproleimagesfrom114humansubjects,capturedin2002.For eachuser, between3and9images weretakenondierent daysandundervaryingposeandlightingconditions. CollectionF: 9423D(depthimages) andcorresponding2Dproleimagesfrom302humansubjects,capturedin2003and2004. CollectionG: 7383D(depthimages)andcorresponding2Dproleimagesfrom235humansubjects,capturedbetween2003and2005 CollectionJ2: 18003D(depthimages)andcorresponding2Dproleimagesfrom415humansubjects,capturedbetween2003and2005[18]. Collection NDO-2007: 7398 3D and corresponding 2D images of 396 human sub-ject faces. The database contains dierent yaw and pitch poses, which are encodedinthelenames[19].2.3 WPUT-DB(a)Goodquality (b) Occlusion byhair(c)Sparselighting (d) Occlusion byjeweleryFigure2:Example images from the WPUT ear database [20]. The database contains earphotographsofvaryingqualityandtakenunderdierentlightingconditions.Furthermorethedatabasecontainsimages,wheretheearisoccludedbyhairorbyearrings.TheWestPommeranianUniversityofTechnologyhascollectedaneardatabasewiththegoal of providingmorerepresentativedatathancomparablecollections (http://ksm.wi.zut.edu.pl/wputedb/) [20]. Thedatabasecontains 501subjects of all agesand2071images intotal. For eachsubject, thedatabasecontains between4and8images,whichweretakenondierentdaysandunderdierentlightingconditions. Thesubjects are also wearing headdresses, earrings and hearing aids, and in addition to this,someearsareoccludedbyhair. InFigure2, someexampleimagesfromthedatabaseareshown. Thepresenceofeachofthesedisruptivefactorsisencodedinthelenamesoftheimages. ThedatabasecanbefreelydownloadedfromthegivenURL.52.4 IITDelhi(a)Example1 (b)Example2 (c)Example3Figure3:ExampleimagesfronIITDelhieardatabase[21].The IIT Delhi Database is provided by the Hong Kong Polytechnic University (http://www4.comp.polyu.edu.hk/~csajaykr/IITD/Database\_Ear.htm) [21]. It containsear images that were collectedbetweenOctober 2006 andJune 2007 at the IndianInstituteofTechnologyDelhi inNewDelhi (seeFigure3). Thedatabasecontains121subjects,andatleast3imagesweretakenpersubjectinanindoorenvironment,whichmeansthatthedatabaseconsistsof421imagesintotal.2.5 IITKanpurThe IITKdatabase was contributedbythe IndianInstitute of TechnologyinKan-pur(http://www.cse.iitk.ac.in/users/biometrics/)[22]. Thisdatabaseconsistsoftwosubsets. SubsetI: Thisdatasetcontains801sidefaceimagescollectedfrom190subjects.Numberofimagesacquiredfromanindividualvariesfrom2to10. Subset II: Theimages inthis subset weretakenfrom89individuals. For eachsubject9imagesweretakenwiththreedierentposes. Eachposewascapturedatthreedierentscales. Mostlikely,allimagesweretakenonthesameday. ItisnotstatedwhethersubsetIIcontainsthesamesubjectsassubsetI.2.6 ScFaceTheSCfacedatabaseisprovidedbytheTechnical Universityof Zagreb(http://www.scface.org/) [23] and contains 4160 images from 130 subjects. The aim of the databaseis toprovide adatabase, whichis suitable for testingalgorithms under surveillancescenarios. Unfortunately, all surveillancecameraimagesweretakenatafrontal angle,such that the ears are not visible on these images. However the database also contains aset of high resolution photographs from each subject, which show the subject at dierentposes. These poses include views of the right and left prole, as shown in Figure 4. Eventhoughthesurveillancecameraimages arelikelytobeunsuitablefor ear recognitionstudies,thehighresolutionphotographscouldbeusedforexaminingresistancetoposevariationsofanalgorithm.6(a)Halfleftprole (b)Fullleftprole (c)FullrightproleFigure4:SCface example images [23]. These images showexamples for the pho-tographed pictures, not for the pictures collected with the surveillance camerasystem.2.7 SheeldFaceDatabaseThis database was formerly known as the UMIST (http://www.sheffield.ac.uk/eee/research/iel/research/face)databaseandconsistsof 564imagesof 20subjectsofmixedraceandgender. Eachsubjectisphotographedinarangeofdierentyawposes,includingafrontalviewandproleviews.2.8 YSUThe Youngston State University collected a new kind of biometric database for evaluationforensic identication systems [24]. For each of the 259 subjects, 10 images are provided.The images are grabbed from a video stream and show the subject in poses between zeroand 90 degrees. This means that the database contains right prole images and a frontalviewimageforeachsubject. Italsocontainshanddrawnsketchesfrom50randomlyselected subjects from a frontal angle. However this part of the database is not of interestforearrecognitionsystems.2.9 NCKUThe National Cheng Kung University in Taiwan has collected an image database, whichconsists of 37 images for each of the 90 subjects. It can be downloaded from the univer-sitys website (http://robotics.csie.ncku.edu.tw/Databases/FaceDetect_PoseEstimate.htm#Our_Database_). Eachsubjectisphotographedindierentanglesbetween-90de-grees (left prole) and 90 degrees (right prole) in 5 degree steps. Figure 5 some examples7(a)40degreesrotation (b)65degreesrotation (c)90degreesrotationFigure5:Some example images from the NKCU face database, showing the same subjectatdierentangles.forthisaredisplayed. Suchaseriesofimagesiscollectedattwodierentdaysforeachofthesubjects. Allimagesweretakenunderthesamelightingconditionsandwiththesamedistancebetweenthesubjectandthecamera.Asthisdatawasoriginallycollectedforfacerecognition,someoftheearsarepartlyor fully occluded by hair, which make this data challenging for ear detection approaches.Consequently,onlyasubsetofthisdatabaseissuitableforearrecognition.2.10 UBEARdatasetThedataset presentedin[25] contains images fromtheleft andtheright ear of 126subjects. Theimages weretakenunder varyinglightingconditions andthesubjectswerenotaskedtoremovehair, jewelryorheaddressesbeforetakingthepictures. Theimagesarecroppedfromvideostream,whichshowsthesubjectindierentposes,suchaslookingtowardsthecamera,upwardsordownwards.Additionally, thegroundtruthfor theears positionis providedtogether withthedatabase, which makes it particularly convenient for researches to study the accuracy ofeardetectionandtostudytheearrecognitionperformanceindependentlyfromanyeardetection.3 EarDetectionThis sectionsummarizes the state of the art inautomatic ear detectionin2Dand3Dimages respectively. Basicallyall ear detectionapproaches arerelyingonmutualpropertiesoftheearsmorphology,liketheoccurrenceofcertaincharacteristicedgesorfrequency patterns. Table 1 gives a short overview of the ear detection methods outlinedbelow. Theupperpartofthetablecontainsalgorithmsfor3Dearlocalization,whereasthelowerpartlistsalgorithmsdesignedforeardetectionin2Dimages.Chen and Bhanu propose three dierent approaches for ear detection. In the approachfrom[26] ChenandBhanutrainaclassier, whichrecognizesaspecicdistributionofshape indices, which are characteristic for the ears surface. However this approach onlyworks on prole images and is sensitive to any kind of rotation, scale and pose variation.8Table1:SummaryofAutomaticeardetectionmethodsfor2Dand3DimagesPublication DetectionMethod Database Perf.#Img TypeChen&Bhanu[28] ShapemodelandICP 700 3D 87.71%Chen&Bhanu[27] HelixShapeModel 213 3D 92.6%Zhouetal.[37] HistogramsofCategorizedShapes 942 3D 100%Prakash&Gupta[32] connectivitygraph 1604 3D 99.38%Abazaetal.[33] Cascadedadaboost 940 2D 88.72%AnsariandGupta[30] Edge detectionandcurvature estima-tion700 2D 93.34%Alastairetal.[39] Raytransform 252 2D 98.4%Alvarezeta.l [36] Ovoidmodel NA 2D NAArbab-Zavar &Nixon[38]HoughTransform 942 2D 91%Arbab-Zavar &Nixon[45]Log-Gabor lters and wavelet transform 252 2D 88.4%Attarchietal.[29] Edgedetectionandlinetracing 308 2D 98.05%Chen&Bhanu[26] Template Matching withShape indexhistograms60 2D 91.5%Islametal.[85] Adaboost 942 2D 99.89%Jeges&Mate[62] Edgeorientationpattern 330 2D 100%Kumaretal.[40] Edgeclusteringandactivecontours 700 2D 94.29%Liu&Liu[86] Adaboostandskincolourltering 50 2D 96%Prakash&Gupta[31] SkinColourandgraphmatching 1780 2D 96.63%Shihetal.[87] Arc-MaskingandAdaBoost 376 2D 100%Yan&Bowyer[18] ConchaDetectionandactivecontours 415 2D >97.6Yuan&Mu[35] CAMSHIFTandaontourtting Video 2D NA9(a)Originalimage,edgeimageandHoughtransform[38](b)Originalimageandraytransform[39](c)Original image, edge enhanced image and correspondingedgeorientationmodel[62]Figure6:ExamplesfordierenteardetectiontechniquesIntheirlatereardetectionapproachesChenandBhanudetectedimageregionswithalargelocal curvaturewithatechniquetheycalledstepedgemagnitude[27]. Thenatemplate, which contains the typical shape of the outer helix and the anti-helix, is ttedtoclustersof lines. In[28] whereChenandBhanunarrowedthenumberof possibleearcandidatesbydetectingtheskinregionrstbeforethehelixtemplatematchingisapplied on the curvature lines. By fusing color and curvature information, the detectionrate could be raised to 99.3% on the UCR dataset and 87.71% on UND collection F andasubsetof collectionG. TheUCRdatasetisnotpubliclyavailableandishencenotcoveredinSection2. Foradescriptionofthisdatasetsee[17].AnotherexampleforeardetectionusingcontourlinesoftheearisdescribedbyAt-trachietal.[29]. Theylocatetheoutercontouroftheearbysearchingforthelongestconnectededgeintheedgeimage. Byselectingthetop,bottom,andleftpointsofthedetected boundary, they form a triangle with the selected points. Further the barycenterofthetriangleiscalculatedandselectedasreferencepointforimagealignment. Ansarietal. alsouseanedgedetectorintherststepoftheirearlocalizationapproach[30].Theedgesareseparatedintotwocategories,namelyconvexandconcave. Convexedgesarechosenascandidatesforrepresentingtheoutercontour. Finallythealgorithmcon-nectsthecurvesegmentsandselectsthegureenclosingthelargestareaforbeingthe10outer ear contour. It shouldbenotedthat theIITKdatabaseandUSTBII alreadycontaincut-outearimages. Henceitcanbeputintoquestion, whetherthedetectionratesof93.34%and98.05%canbereproducedunderrealisticconditions.A recent approach on 2D ear detection using edges is described by Prakash and Guptain[31]. Theycombineskinsegmentationandcategorizationof edgesintoconvexandconcaveedges. Afterwardstheedgesintheskinregionaredecomposedintoedgeseg-ments. These segments are composed to form an edge connectivity graph. Based on thisgraph the convex hull of all edges,which are believed to belong to the ear,is computed.Theenclosedregionisthenlabeledastheearregion. Incontrastto[29], PrakashandGupta prove the feasibility of edge-based ear detection on full prole images, where theyachieved a detection rate of 96.63% on a subset of the UND-J2 collection. In [32] proposethesameedgeconnectivityforearrecognitionon3Dimages. Insteadofedges,theyusediscontinuitiesinthedepthmapforextractingtheinitialedgeimageandthenextracttheconnectivitygraph. Intheirexperiments, theyusethe3Drepresentationsof thesamesubsetasin[31]andreportadetectionrateof99.38%. Moreovertheyshowthatthedetectionrateoftheirgraph-basedapproachisnotinuencebyrotationandscale.Jedges and Mate propose another edge-based ear detection approach, which is likely tobe inspired by ngerprint recognition techniques. They train a classier with orientationpattern, whichwerepreviouslycomputedfromearimages. Likeothernaiveclassiers,their methodis not robust against rotationandscale. Additionallythe classier islikelytofail underlargeposevariations, becausethiswill aecttheappearanceoftheorientationpattern.Abazaetal.[33] andIslametal.[34] useweakclassiersbasedonHaar-waveletsinconnection with AdaBoost for ear localization. According to Islam etal., the training ofthe classier takes several days, however once the classier is set up, ear detection is fastand eective. Abaza et al. use a modied version of AdaBoost and report a signicantlyshortertrainingphase. Theeectivenessoftheirapproachisprovedinevaluationsonvedierentdatabases. Theyalsoincludesomeexamplesof successful detectionsonimages from the internet. As long as the subjects pose does not change, weak classiersaresuitableforimageswhichcontainmorethanonesubject. DependingonthetestsetAbazaet al. achievedadetectionratebetween84%and98.7%ontheSheeldFacedatabase. Onaverage,theirapproachsuccessfullydetected95%ofallears.YanandBrowyerdevelopedaneardetectionmethodwhichfusesrangeimagesandcorresponding2Dcolorimages[18]. Theiralgorithmstartsbylocatingtheconchaandthenusesactivecontoursfordeterminingtheearsouterboundary. Theconchaservesasthereferencepointforplacingthestartingshapeoftheactivecontourmodel. Eventhoughtheconchaiseasytolocalizeinproleimages, itmaybeoccludediftheheadpose changes or if a subject is wearing a hearing aid or ear phones. In their experimentsYan and Browyer only use ear images with minor occlusions where the concha is visible;henceit couldneither beprovednor disprovedwhether their approachis capableofreliablydetectingearsiftheconchaisoccluded.YuanandMudevelopedamethodforreal-timeeartrackinginvideosequencesbyapplying Continuously Adaptive Mean Shift (CAMSHIFT) to video sequences [35]. TheCAMSHIFTalgorithmisfrequentlyusedinfacetrackingapplicationsandisbasedon11region matching and a skin color model. For precise ear segmentation, the contour ttingmethodbasedonmodiedactiveshapemodels, whichhavebeenproposedbyAlvarezet al. isapplied[36]. YuanandMureportadetectionrateof100%, howeverthetestdatabase only consisted of two subjects. Nevertheless their approach appears to be verypromising for surveillance applications but needs to be further evaluated in more realistictestscenarios.Shihetal. determineearcandidatesbylocalizingarc-shapededgesinanedgeimage.Subsequently the arc-shaped ear candidates are veried by using an Adaboost classier.Theyreportadetectionrate100%onadataset, whichconsistsof376imagesfrom94subjects.Zhouet al. traina3Dshapemodel inorder torecognizethehistogramof shapeindexesof thetypical ear[37]. Similarlytotheapproachesof Abazaet al. andIslamet al., aslidingwindowof dierentsizesismovedovertheimage. TheeardescriptorproposedbyZhouet al. is built fromconcatenatedshape indexhistograms, whichareextractedfromsub-blocksinsidethedetectionwindow. Fortheactualdetection, aSupport Vector Machine (SVM) classier is trained to decide whether an image region istheearregionornot. Asfarasweknow,thisisthersteardetectionapproach,whichdoesnotrequirehavingcorrespondingtextureimagesinadditiontotherangeimage.Zhouet al. evaluatedtheirapproachonimagesfromtheUNDcollectionsandreportadetectionrateof 100%. Itshouldbenotedthatthisapproachwasnottestedunderrotation, posevariationsandmajorocclusions, butundertheimpressionof thegoodperformance,wethinkthisisaninterestingtaskforfutureresearch.Eardetectionmethodsbasedonimagetransformationshavetheadvantageofbeingrobustagainstout-of-planerotations. Theyaredesignedtohighlightspecicpropertiesoftheouterear,whichoccurineachimagewheretheearisvisiblenomatterinwhichposetheearhasbeenphotographed. In[38]theHoughtransformisusedforenhancingregionswithahighdensityof edges. Inheadproleimages, ahighdensityof edgesespeciallyoccurs inthe ear region(see Figure 6(a)). In[38] it is reportedthat theHoughtransformbasedeardetectiongetstrappedwhenpeoplewearglassessincetheframeintroducesadditional edgestotheimage. Thisespeciallyoccursintheeyeandnose region. The ear detection approach based on Hough transform was evaluated on theimagesintheXM2VTSdatabase(see[17]foradetaileddatabasedescription),whereadetectionrateof91%wasachieved.Theraytransformapproachproposedin[39]isdesignedtodetecttheearindierentposes andtoignorestraight edges intheimage, whichcanbeintroducedbyglassesor hair. Raytransformuses alight rayanalogytoscanthe image for tubular andcurvedstructuresliketheouterhelix. Thesimulatedrayisreectedinbrighttubularregionsandhencetheseregionsarehighlightedinthetransformedimage. Sinceglassframeshavestraightedges, theyarenothighlightedbytheraytransform(seeFigure6(b)). Usingthis methodAlastair et al. achievedamimpressiverecognitionrateof98.4%ontheXM2VTSdatabase. Hence,theraytransformapproachbyAlastairetal.outperformsHoughtransform, mostlikelybecauseitismorerobustagainstdisruptivefactorssuchasglassesorhair.Arecentapproachfor2Deardetectionisdescribedin[40]. Kumaretal. proposeto12extractearsfrom2Dimagesbyusingedgeimagesandactivecontours. Theyevaluatetheirapproachonadatabase,whichconsistsof100subjectswith7imagespersubject.Aspecialimagingdevicewasusedforcollectingthedata. Thisdevicemakessurethatthedistancetothecameraisconstantandthatthelightingconditionsarethesameforallimages. Withinthissettingadetectionrateof94.29%isreported.4 2DEarRecognitionEachearrecognitionsystemconsistsofafeatureextractionandafeaturevectorcom-parisonstep. Inthis surveywedivideear recognitionapproaches intofour dierentsubclasses namely holistic approaches, local approaches, hybrid approaches and statisti-calapproaches.IntheTables2and3all2Dearrecognitionapproachesmentionedinthispaperaresummarizedinchronologicalorder.4.1 HolisticDescriptors(a) Concentric Circles[57](b)SIFTfeatures[53] (c)ActiveContour[62] (d)ForceField[41]Figure7:Examplesforfeatureextractionfor2Dearimages.Anotherapproach, whichhasgainedsomepopularityistheForceFieldTransformbyHurley[41]. TheForceFieldtransformationapproachassumesthatpixelshaveamutual attractionproportional totheir intensities andinverselytothesquareof thedistance between them rather like Newtons universal law of gravitation. The associatedenergyeldtakestheformofasmoothsurfacewithanumberofpeaksjoinedbyridges(seeFigure 7(d)). Usingthis method, Hurleyet al. achievedarank-1performanceof morethan99%ontheXM2VTSdatabase(252images). Buildingontheseresults,Abdel-Mottaleb and Zhou use a 3D representation of the force eld for extracting pointslyingonthepeakof the3Dforceeld[42]. Becausetheforceeldconvergedattheoutline of the ear, the peaks in the 3D representation basically represent the ear contour.Nonetheless, the force eld method is more robust against noise than other edge detector,13Table2:Summaryof approaches for 2Dear recognitionapproaches, part 1. Unlessstateddierently,performancealwaysreferstorank-1performance.Publication Summary Database Perf.#Subj #ImgBurge and Burger[88]VornoiDistanceGraphs NA NA NAYuanandMu[66] FullSpaceLDAwithOuterHelixFeaturePoints79 1501 86.76%Hurley[41] ForceFieldTransform 63 252 99%Morenoetal.[89] Geometric features with Compres-sionNetwork28 268 93%Yuizonoetal.[73] GeneticLocalSearch 110 660 99%Victoretal.[67] PCA 294 808 40%Changetal.[68] PCA 114 464 72.7%Abdel-MottalebandZhou[42]ModiedForceFieldTransform 29 58 87.9%Muetal.[58] Geometrical measures on edge im-ages77 308 85%Abateetal.[46] GeneralFourierDescriptor 70 210 88%Luetal.[65] ActiveShapeModelandPCA 56 560 93.3%Yuanetal.[90] Non-Negative Matrix Factoriza-tion77 308 91%Arbab-Zavar et al.[53]SIFTpointsfromearmodel 63 252 91.5%Jedges and Mate[62]DistortedEarModel withfeaturepoints28 4060 5.6%EERLiuetal.[63] Edge-based features from dierentviews60 600 97.6%Nanni and Lumini[72]GaborFiltersandSFFS 114 464 80%Rahmanetal.[59] GeometricFeatures 100 350 87%Sanaetal.[47] Haar Wavelets and Hamming Dis-tance600 1800 98.4%Arbab-Zavar andNixon[45]Log-GaborFilters 63 252 85.7%Choras[57] Geometryofearoutline 188 376 86.2%14Table3:Summaryof approaches for 2Dear recognitionapproaches, part 2. Unlessstateddierently,performancealwaysreferstorank-1performance.Publication Summary Database Perf.#Subj #ImgDongandMu[43] ForceFieldTransformandNKFDA29 711 75.3%GuoandXu[60] Local Binary PatternandCNN77 308 93.3%Nasseemetal.[76] Sparserepresentation 32 192 96.88%Wangetal.[91] Haar Wavelets and Lo-calBinaryPatterns79 395 92.41%XieandMu[71] Locally Linear Embed-ding79 1501 80%Yaqubietal.[74] HMAXandSVM 60 180 96.5%ZhangandMu[70] Geometrical Features,ICA and PCA withSVM77 308 92.21Badrinath andGupta[54]SIFTlandmarks fromearmodel106 1060 95.32%Kiskuetal.[55] SIFT from dierentColorSegments400 800 96.93%Wang and Yuan[50]Low-Order MomentInvariants77 308 100%Alarajetal..[69] PCAwithMLFFNNs 17 85 96%Bustardetal.[52] SIFTPointMatches 63 252 96%De Marisco et al.[48]Partitioned IteratedFunction System(PIFS)114 228 61%Gutierrez etal. [75] MNN with SugenoMeasuresandSCG77 308 97%Wangetal.[49] Moment Invariantsand BP Neural Net-workNA 60 91.8%Wang and Yuan[44]Gabor Wavelets andGDA77 308 99.1%Fooprateepsiri andKurutach[92]Trace and FourierTransform68 68 97%Prakash and Gupta[22]SURFandNNclassi-er300 2066 2.25%EERKumaretal.[40] SIFT 100 700 95%GAR,0.1%FARWangandYan[61] Local Binary Patternand Wavelet Trans-form77 308 100%Kumar and Wu [21] Phase encoding withLogGaborlters221 753 95.93%15such as Sobel or Canny. Using this approach, Abdel-Mottaleb and Zhou achieved a rank-1performanceof87.93%onadatasetwithconsistsof103earimagesfrom29subjects.DongandMu[43] addposeinvariancetotheedges, whichareextractedbyusingthe force eldmethod. This is achievedwithnull space kernel shier discriminantanalysis(NKFDA),whichhasthepropertyofrepresentingnon-linearrelationsbetweentwodatasets. DongandMuconductedexperimentsontheUSTBIVdataset. Beforefeatureextraction, theearregionwascroppedoutmanuallyfromtheimagesandtheposeisnormalized. Forposevariationof 30theyreportarank-1reconditionrateof72.2%. Forposevariationsof45therank-1performancedroppedto48.1%.In a recent publication of Kumar and Wu [21] they present an ear recognition approach,whichusesthephaseinformationof Log-Gaborltersforencodingthelocal structureof theear. Theencodedphaseinformationisstoredinnormalizedgreylevel images.Intheexperiments, theLog-Gabor approachoutperformedforceeldfeatures andalandmark-basedfeatureextractionapproach. Moreover, dierentcombinationsofLog-Gabor lters were compared with each other. The rank-e performance for the Log-Gaborapproaches ranges between 92.06% and 95.93% on a database which contains 753 imagesfrom221subjects.Therichstructureoftheouterearresultsinspecictextureinformation, whichcanbemeasuredusingGaborlters. WangandYuan[44] extractlocal frequencyfeaturesby using a battery of Gabor lters and then select the most distinctive features by usinggeneral discriminant analysis. Intheir experiments onthe USTBII database, theycomparedtheperformanceimpactof dierentsettingsfortheGaborlters. Dierentcombinationsof orientationandscalesintheltersetsarecomparedwitheachotheranditwasfoundthatneitherthenumberofscalesnorthenumberoforientationshasamajor impact ontherank-1performance. Thetotal rank-1performanceof WangandYuans approachis 99.1%. Ina similar approachArbab-Zavar andNixon[45]measuredtheperformanceofGaborltersintheXM2VTSdatabasewheretheyreportarank-1performanceof 91.5%. Acloser lookat theGabor lter respondedshowedthatthefeaturevectorsarecorruptedbyocclusionorotherdisruptivefactors. Inordertoovercomethis, morerobust comparisonmethodis proposed, whichresultedinanimprovedrecognitionrateof97.4%.Abate et al. [46] use a generic Fourier descriptor for rotationandscale invariantfeaturerepresentation. Theimageistransformedintoapolarcoordinatesystemandthentransformedintofrequencyspace. Inorder tomake sure, that the centroidofthepolarcoordinatesystemisalwaysatthesameposition, theearimageshavetobealignedbeforetheycanbetransformedintothepolarcoordinatesystem. Theconchaserves as a reference point for the alignment step, such that the center point of the polarcoordinate system is always located in the concha region. The approach was tested on aproprietarydataset,whichcontains282earimagesintotal. Theimagesweretakenontwodierentdaysandindierentroll andyawposes. Therank-1performanceof theFourierdescriptorvariesdependingontheposeangle. For0posevariationtherank-1performanceis96%, butifdierentposesareincludedintheexperiments, itdropsto44%for15and 19%for30.Intheworkof Fooprateepsiri andKurutachexploittheconceptsof multi-resolution16Trace transform and Fourier transform. The input images from the CMU PIEdatabaseare serialized by using the trace transform and stored in a feature vector. The advantageofthetracetransformisthattheresultingfeaturevectorisinvarianttorotationsandscale. Furthermore Fooprateepsiri and Kurutach show that their descriptor is also robustagainstposevariations. Intotaltheyreportarank-1performanceof 97%.Sanaetal. useselectedwaveletcoecientsextractedduringHaar-Waveletcompres-sionfor featurerepresentation[47]. Whileapplyingthefour level wavelet transformseveral timesontheear image, for eachiterationtheystoreoneof thederivedcoef-cientsinafeaturevector. Thereportedaccuracyof theiralgorithmis96%andwasachievedonthebasisoftheIITKdatabaseandontheSaugordatabase(350subjects).AfeatureextractionsystemcalledPIFSisproposedbyDeMariscoetal.[48]. PIFSmeasures the self-similarity in an image by calculating ane translations between similarsubregionsofanimage. Inordertomaketheirsystemrobusttoocclusion,DeMariscoet al. dividedtheearimageintoequallylargetiles. If onetileisoccluded, theothertilesstillcontainasucientlydistinctivesetoffeatures. DeMariscoetal. couldshowthattheirapproachissuperiortootherfeatureextractionmethodsunderthepresenceof occlusion. The experiments of De Mariscoet al. have beenconductedinordertoassessthesystemperformanceindierentocclusionscenarios. ThebasisforthesetestswastheUNDcollectionEandtherst100subjectsof theFERETdatabase. Ifocclusion occurs on the reference image, a rank-1 performance of 61% (compared to 40%onaveragewithotherfeatureextractionmethods)isreported. Withoutocclusion, therank-1performanceis 93%.Moment invariants are astatistical measure for describingspecic properties of ashape. Wanget al. [49] composesixdierentfeaturevectorsbyusingsevenmomentinvariants. They also show that each of the moment invariants is robust against changesinscaleandrotation. Thefeaturevectorsareusedastheinputforabackpropagationneural network which is trained to classify the moment invariant feature sets. Based on aproprietary database of 60 ear images, they report a rank-1 performance of 91.8%. In [50]WangandYuancomparethedistinctivenessofdierentfeatureextractionmethodsontheUSTBI database. Theycomparetherank-1performanceof Fourier descriptors,Gabor-Transform, MomentInvariantsandstatistical featuresandcometotheconclu-sionthatthehighestrecognitionratecanbeachievedbyusingmomentinvariantsandGabor transform. For both feature extractionmethods Wang and Yuan report a rank-1performanceor100%.4.2 Local DescriptorsScaleinvariantFeatureTransform(SIFT)isknowntobearobustwayforlandmarkextractioneveninimageswithsmallposevariationsandvaryingbrightnessconditions[51]. SIFTlandmarkscontainameasureforlocalorientation;theycanalsobeusedforestimatingtherotationandtranslationbetweentwonormalizedearimages. Bustardetal. showed that SIFT can handle pose variations up to 20 degrees [52]. However it is nota trivial to assign a SIFT landmark with its exact counterpart, especially in the presenceof posevariations. Inhighlystructuredimageregions, thedensityandredundancyof17SIFTlandmarksissohigh,thatexactassignmentisnotpossible. Hencethelandmarkshavetobelteredbeforetheactual comparisoncanstart. Arbab-Zavaret al. [53] aswell as Badrinath and Gupta [54] therefore train a reference landmark model, which onlycontainsasmallnumberofnon-redundantlandmarks. ThislandmarkmodelisusedforlteringtheSIFTlandmarks, whichwereinitiallydetectedintheprobeandreferenceear. Havingthelteredlandmarksitispossibletoassigneachof thelandmarkswithitsmatchingcounterpart. Figure7(b)showsanexampleforSIFTlandmarksextractedfromanearimages,whichwereusedastrainingdataforthereferencelandmarkmodelintheworkofArbab-Zavaretal.. BecauseArbab-Zavaretal. alsousedtheXM2VTSdatabase for evaluation, their results can be directly compared to the rank-1 performancereportedbyBustardandNixon. Arbab-Zavaretal. achievedarank-1performanceof91.5%. WiththemorerecentapproachbyBustardandNixontheperformancecouldbeimprovedto96%. UsingtheIITDelhidatabaseKumaretal. reportaGARof95%andaFARof0.1%whenusingSIFTfeaturepoints.Kiskuet al. address theproblemof correct landmarkassignment bydecomposingtheearimageintodierent colorsegments[55]. SIFTlandmarksareextractedfromeachsegmentseparately, whichreducesthechanceof assigningSIFTlandmarksthatarenot representingthesamefeatures. Usingthis approach, Kiskuet al. achievearank-1performanceof96.93%.Arecent approachbyPrakashandGupta[22] fuses SpeededUpRobust Features(SURF) [56] featurepointsfromdierent imagesof thesamesubject. Theyproposetouseseveral inputimagesforenrolmentandtostoreall SURFfeaturepointsinthefusedfeaturevector, whichcouldbefoundintheinputimages. Thesefeaturesetsarethenusedfortraininganearestneighborclassierforassigningtwocorrelatedfeaturepoints. If the distance between two SURF feature points is less than a trained threshold,theyareconsideredtobecorrelated. Theevaluationof thisapproachwascarriedouton the UND collection E and the two subsets of the IIT Kanpur database. Prakash andGupta tested the inuence of dierent parameters for SURF features and for the nearestneighbor classier. Dependingonthecompositionof theparameters theEERvariesbetween6.72%and2.25%.Choras proposes a set of geometric feature extraction methods inspired by the work ofIannarelli[57]. Heproposesfourdierentwaysoffeaturelocationinedgeimages. Theconcentric circles method uses the concha as reference points for a number of concentriccircleswithpredenedradii. Theintersectionpointsofthecirclesandtheearcontoursareusedasfeaturepoints(seeFigure7(a).). Anextensionofthisisthecontourtracingmethods, whichuses bifurcations, endpoints andintersectingpoints betweentheearcontours as additional features. Intheanglerepresentationapproach, Choras drawsconcentriccirclesaroundeachcenterpointofanedgeandusestheanglesbetweenthecenter point andthe concentric circles intersectingpoints for feature representation.Finally the triangle ratio method determines the normalized distances between referencepoints and uses them for ear description. Choras conducted studies on dierent databaseswhere he reported recognition rates between 86.2% and 100% on a small database o 12subjectsandafalserejectratebetween0%and9.6%onalargerdatabasewith102earimages.18Similarapproacheswhichareusingtheaspectratiobetweenreferencepointsontheear contours are proposedbyMuet al. witharank-1performance of 85%ontheUSTB II database [58] and Rahman etal. [59]. Rahman etal. evaluated their approachonadatabase, whichconsistsof 350imagesfrom100subjects. Theyreportarank-1 performance of 90%. For images, whichwere takenondierent days the rank-1performancedroppedto88%.Localbinarypatterns(LBP)areatechniqueforfeatureextractiononthepixellevel.LBPencodethelocal neighborhoodof apixel bystoringthedierencebetweentheexaminedpixelanditsneighbors. Guoetal.[60]extractLBPfromtherawearimagesand create histograms describing the distribution of the local LBP. Then a cellular neuralnetworkoftrainedtodistinguishbetweentheLBPofdierentsubjectsintheUSTBIIdatabase.InthebyWangandYan[61]thedimensionalityofthefeaturevectorisreducedwithlinear discriminant analysis before a Euclidean distance measure quanties the similarityof two feature vectors. Wang and Yan evaluated their approach on the USTB II datasetandreportarank-1performanceof100%.4.3 HybridApproachesTheapproachofJedgesandMateistwofold[62]. Inarstfeatureextractionsteptheygenerate an average edge model from a set of training images. These edges represent theouterhelixcontouraswell asthecontoursof theantihelix, thefossatriangularisandthe concha. Subsequently each image is enrolled by deforming the ear model until it tstheactualedgesdisplayedintheprobeearimage. Thedeformationparameters,whichwere necessary for the transformation, are the rst part of the feature vector. The featurevectoriscompletedbyaddingadditionalfeaturepointslyingofintersectionsbetweenapredenedsetof axesandthetransformedmainedges. Theaxesdescribetheuniqueoutline of ear. Figure 7(c) shows the edge enhanced images with tted contours togetherwiththeadditional axesforreferencepointextraction. TheyreportanEERof 5.6%usingadatabasewithcroppedimagesandwithoutposevariations.Liuet al. combinefrontandbacksideviewof theearbyextractingfeaturesusingthe triangle ratio method and Tchebichef moment descriptors [63]. Tchebichef momentsareasetoforthogonalmomentfunctionsbasedondiscreteTchebichefpolynomialsandhavebeenintroducedasamethodforfeaturerepresentationin2001[64]. Thebacksideof theearisdescribedbyanumberof linesthatareperpendiculartothelongestaxisintheearcontour. Theselinesmeasurethelocal diameteroftheauricleatpredenedpoints. Therank-1performanceofthiscombinedapproachisreportedtobe97.5%. Ifonlythefrontviewisused,therank-1performanceis95%andforthebacksideimages,Liuetal. report86.3%rank-1performance.Luetal.[65]aswellasYuanandMu[66]usetheactiveshapemodelforextractingtheoutlineof theear. Luet al. areusingmanuallycroppedearimagesfrom56sub-jectsindierentposes. Afeatureextractorstoresselectedpointsontheoutlineoftheeartogetherwiththeirdistancetothetragus. Beforeapplyingalinearclassier, thedimensionality of the feature vectors is reduced by principal component analysis (PCA).19Luetal. comparetherank-1performanceofpipelineswhereonlytheleftortherightearwasusedforidenticationandalsoshowthatusingbothearsincreasestherank-1performancefrom93.3%to95.1%. IntheUSTBIIIdatabaseYuanandMureportarank-1performanceof 90%istheheadrotationislowerthan15. Forrotationanglesbetween20and60therank-1performancedropsto80%.4.4 ClassiersandStatistical ApproachesVictoret al. weretherstresearchgrouptotransfertheideaof usingtheeigenspacefromfacerecognitiontoearrecognition[67]. Theyreportedthattheperformanceoftheearasafeatureisinferiortotheface. ThismaybeduetothefactthatintheirexperimentsVictoretal. consideredtheleftandtherighteartobesymmetric. Theyusedtheoneearfortrainingandtheotherearfortesting, whichcouldhaveloweredtheperformanceofPCAinthiscase. Thereportedrank-1performanceis40%. Witharank-1performance of 72.2%inthe UNDcollectionE, Changet al. [68] report asignicantlybetterperformancethanVictoretal.. Alarajetal.[69]publishedanotherstudy, where PCA is used for feature representation in ear recognition. In their approacha multilayer feed forward neural network was trained for classication of the PCA basedfeaturecomponents. Theobservedarank-1performanceof 96%, andhenceimprovedthepreviousresultsbyVictoret al. andChanget al.. HoweveritshouldbenoticedthatthisresultisonlybasedonasubsetofoneoftheUNDcollections,whichconsistsof85earimagesfrom17subjects.ZhangandMuconductedstudiesontheeectivenessofstatistical methodsincom-binationwithclassiers. In[70]theyshowthatindependentcomponentanalysis(ICA)ismoreeectiveontheUSTBIdatabasethanPCA.TheyrstusedPCAandICAforreducing the dimensionality of the input images and then trained an SVM for classifyingtheextractedfeaturevectors. Furthermoretheinuenceof dierenttrainingsetsizesontheperformancewasmeasured. Dependingonthesiteofthetrainingsettherank-1performanceforPCAvariesbetween85%and94.12%,whereastherank-1performanceforICAvariesbetween91.67%and100%.XieandMu[71] proposeanimprovedlocallylinearembedding(LLE)algorithmforreducingthedimensionalityof ear features. LLEis atechniquefor projectinghigh-dimensional data points into a lower dimensional coordinate system while preserving therelationshipbetweenthesingledatapoints. Thisrequiresthedatapointstobelabeledinsomeway, sothattheirrelationshipisxed. TheimprovedversionomLLEbyXieandMueliminatedtheproblembyusingadierentdistancefunction. FurtherXieandMu show, that LLE is superior to PCA and Kernel PCA, if the input data contains posevariations. TheirstudieswereconductedontheUSTBIII databaseshowedthattherank-1performanceofregularLLE(43%)isimprovedsignicantlybytheirmethodto60.75%. Iftheposevariationisonly10,theimprovedLLEapproachachievedarank-1performanceof90%.IntheirapproachNanniandLumini[72]proposetouseSequentialForwardFloatingSelection(SFFS),whichisastatisticaliterativemethodforfeatureselectioninpatternrecognitiontasks. SFFStriestondthebestsetofclassiersbycreatingasetofrules,20whichbesttsthecurrentfeatureset. Thesetsarecreatedbyaddingoneclassieratatimeandevaluatingitsdiscriminativepowerwithapredenedtnessfunction. Ifthenewset of rulesoutperformsthepreviousversion, thenewruleisaddedtothenalsetofrules. TheexperimentswerecarriedoutontheUNDcollectionEandthesingleclassiers are fused by using the weighted sum rule. SFFS selects the most discriminativesub-windowswhichcorrespondtothettestsetof rules. Nanni andLumini reportarank-1recognitionrateof 80%andarank-5recognitionrateof 93%. TheEERvariesbetween 6.07% and 4.05% depending on the number of sub-windows used for recognition.Yiuzonoet al. considertheproblemof ndingcorrespondingfeaturesinearimagesasanoptimizationproblemandapplygeneticlocalsearchforsolvingititeratively[73].Theyselectlocal subwindowswithvaryingsizeasthebasisforthegeneticselection.In[73]Yiuzonoetal. presentelaboratedresults,whichdescribethebehaviorofgeneticlocal search under dierent parameters, such as dierent selection methods and dierentnumbersofchromosomes. Onadatabaseof110subjectstheyreportarecognitionrateof100%.Yaqubi et al. usefeaturesobtainedbyacombinationof positionandscale-tolerantedgedetectorsovermultiplepositionsandorientationsoftheimage[74]. Thisfeatureextraction method is called HMAX model and is inspired by the visual cortex of primatesand combines simple features to more complex semantic entities. The extracted featuresareclassiedwithanSVNandakNN. Therank-1performanceonasmall datasetof180 cropped ear images from 6 subjects varies between 62% and 100% depending on thekindofbasisfeatures.Morenoet al. implementafeatureextractor, whichlocatessevenlandmarksontheearimage, whichcorrespondtothesalientpointsfromtheworkof Iannarelli. Addi-tionallytheyobtainamorphologyvector, whichdescribestheearasawhole. Thesetwo features are used as the input for dierent neural network classiers. They comparetheperformanceofeachofthesinglefeatureextractiontechniqueswithdierentfusionmethods. Theproprietarytest databaseis composedof manuallycroppedears from168from28subjects. Thebestresultof93%rank-1performancewasmeasuresusingacompressionnetwork. Othercongurationsyieldederrorratesbetween16%and57%.Gutierrez etal. [75] divide the cropped ear images into three equally sized parts. Theupperpartshowsthehelix,themiddlepartshowstheconchaandthelowerpartshowsthelobule. Eachofthesesubimagesisdecomposedbywavelettransformandthenfedinto a modular neural network. In each module of the network a dierent integrators andlearning functions was used. The results of each of the modules are fused in the last stepforobtainingthenaldecision. Dependingonthecombinationbetweenintegratorandlearningfunction, theresultsvarybetween88.4%and97.47%rank-1performanceonthe USTB I database. The highest rank-1 performance is achieved with Sugeno measureandconjugategradient.In[76]Nasseemetal. proposeageneralclassicationalgorithmbasedonthetheoryofcompressivesensing. Theyassumethatmostsignalsarecompressibleinnatureandthatanycompressionfunctionresultsinasparserepresentationofthissignal. IntheirexperimentsintheUNDdatabaseandtheFEUDdatabase, Nasseemetal. showthattheir sparse representation method is robust against pose variations and varying lighting21Table4:Summaryof approachesfor3Dearrecognition. Performance(Perf.) alwaysreferstorank-1performance.Publication ComparisonMethod Database Perf.#Subj #ImgCadavidetal.[84] ICPandShapefromshad-ing462 NA 95%Chen and Bannu[28]LocalSurfacePatch 302 604 96.36%Chen and Bhanu[27]ICPContourMatching 52 213 93.3%Liuetal.[63] Mesh PCA with neural net-work60 600 NALiuandZhang[81] SliceCurveMatching 50 200 94.5%Islametal.[82] ICPwithreducedmeshes 415 830 93.98%Islametal.[93] Local Surface Features withICP-Matching415 830 93.5%Passalisetal.[80] Reference ear model withmorphing525 1031 94.4%Yan and Bowyer[77]ICPusingvoxels 369 738 97.3%Yan and Bowyer[18]ICPusingModelPoints 415 1386 97.8%Zhengetal.[83] LocalBinaryPatters 415 830 96.39%Zhouet. al.[79] Surface Patch Histogramandvoxelization415 830 98.6%,1.6%EERconditions. Therank-1performancevariedbetween89.13%and97.83%, dependingonthedatasetusedintheexperiment.5 3DEarRecognitionIn2Dear recognitionposevariationandvariationincameraposition, so-calledout-of-plane-rotations,arestillunsolvedchallenges. Apossiblesolutionisusing3Dmodelsinstead of photos as references, because a 3D representation of the subject can be adaptedtoanyrotation, scaleandtranslation. Inadditiontothat, thedepthinformationcon-tained in 3D models can be used for enhancing the accuracy of an ear recognition system.However, most 3D ear recognition systems tend to be computationally expensive. In Ta-ble4all3Dearrecognitionsystemsdescribedinthissectionaresummarized.AlthoughICPis originallydesignedtobeanapproachfor imageregistration, the22Figure8:Examplesforsurfacefeaturesin3Dearimages. Theupperimageshowsanexample for ICP-based comparison as proposed in [27], whereas the lower gureillustratesfeatureextractionfromvoxelsasdescribedin[77].registrationerrorcanalsobeusedasameasureforthedissimilarityoftwo3Dimages.BecauseICPisdesignedtobearegistrationalgorithm,itisrobustagainstallkindsoftranslationorrotations. HoweverICPtendstostoptooearly, becauseitgetsstuckinlocalminima. ThereforeICPrequiresthetwomodelstobecoarselypre-alignedbeforene alignment using ICPcan be performed. Chenand Bhanuextract pointclouds fromthecontouroftheouterhelixandtheregisterthesepointswiththereferencemodelbyusingICP[27]. InalaterapproachChenandBhanuuselocal surfacepatches(LSP)instead of points lying on the outer helix [28]. As the LSP consist or less points than theouterhelix, thisreducestheprocessingtimewhileenhancingtherank-1performancefrom93.3%withtheouterhelixpointsto96.63%withLSP.YanandBrowyerdecomposetheearmodel intovoxelsandextractsurfacefeaturesfrom each of these voxels. For speeding up the alignment process,each voxel is assignedan index in such a way that ICP only needs to align voxel pairs with the same index [77](seeFigure8). In[18] YanandBrowyerproposetheusageof pointcloudsfor3Dearrecognition. Incontrastto[27] all pointsof thesegmentedearmodel areused. Thereportedperformancemeasures of 97.3%in[77] and97.8%in[18] is similar but notdirectlycomparable,becausedierentdatasetswereusedforevaluation.Cadavidet al. proposeareal-timeear recognitionsystem, whichreconstructs 3Dmodelsfrom2DCCTVimagesusingtheshapefromshadingtechnique[78]. Thereafterthe3Dmodelincomparedtothereference3Dimages, whicharestoredinthegallery.Model alignment as well as the computation of the dissimilarity measure is done by ICP.Cadavidet al. reportarecognitionrateof 95%onadatabaseof 402subjects. Itisstatedin[78] thattheapproachhasdicultieswithposevariations. In[79] Zhouet.al. useacombinationof local histogramfeaturesvoxel-models. Zhouet. al. reportthattheirapproachisfasterandwithanEERof1.6%itisalsomoreaccuratethantheICP-basedcomparisonalgorithmsproposedbyChenandBhanuandYanandBrowyer.SimlarlytoCadavidetal., Liuetal. reconstruct3Dearmodelsfrom2Dviews[63].Basedonthetwoimagesof astereovisioncamera, a3Drepresentationof theearisderived. Subsequently the resulting 3D meshes serve as the input for PCA. However Liu23etal. donotprovideanyresultsconcerningtheaccuracyoftheirsystembutsincetheydidnotpublishanyfurtherresultsontheirPCAmeshapproach,itseemsthatitisnolongerpursued.Passalisetal. goadierentwayforcomparing3Dearmodelsinordertomakecom-parisonsuitableforareal-timesystem[80]. Theycomputeareferenceearmodelwhichis representativefor theaveragehumanear. Duringenrolment, all referencemodelsaredeformeduntiltheytthereferenceearmodel. Alltranslationsanddeformations,whichwerenecessarytottheeartothereferencemodel arethenstoredasfeatures.If aprobeforauthenticationisgiventothesystem, themodel isalsoadaptedtotheannotatedearmodel inordertogetthedeformationdata. Subsequentlythedeforma-tiondataisusedtosearchforanassociatedreferencemodelinthegallery. Incontrasttothepreviouslydescribedsystems, onlyonedeformationhastobecomputedperau-thenticationattempt. Allotherdeformationmodelscanbecomputedbeforetheactualidenticationprocessisstarted. Thisapproachisreportedtobesuitableforreal-timerecognitionsystems,becauseittakeslessthan1msecforcomparingtwoeartemplates.The increased computing speed is achieved by lowering the complexity class from O(n)2forICP-basedapproachestoO(n)fortheirapproach. Therank-1recognitionrateisreportedtobe94.4%. Theevaluationisbasedonnon-publicdata,whichwascollectedusingdierentsensors.Heng and Zhang propose a feature extraction algorithm based on slice curve compari-son,whichisinspiredbytheprinciplesofcomputertomography[81]. Intheirapproachthe3Dear model is decomposedintoslices alongtheorthogonal axis of thelongestdistancebetweenthelobuleandtheuppermostpartofthehelix. Thecurvatureinfor-mationextractedfromeachsliceisstoredinafeaturevectortogetherwithanindexvalue indicating the slices former position in the 3D model. For comparison the longestcommonsequencebetweentwoslicecurveswithsimilarindexesisdetermined. Theirapproachisonlyevaluatedonanon-publicdataset, whichconsistsof200imagesfrom50subjects. Noinformationabout posevariations or occlusionduringthecapturingexperiment is given. HengandZhangreport arank-1performanceis 94.5%for theidenticationexperimentand4.6%EERforthevericationexperiment.Islam et al. reconnect point clouds describing 3D ear models to meshes and iterativelyreducethenumberoffacesinthemesh[82]. ThesesimpliedmeshesarethenalignedwitheachotherusingICPandthealignmenterrorisusedasthesimilaritymeasureforthetwosimpliedmeshes. InalaterapproachIslametal. extractlocalsurfacepatchesas shown in Figure 9 and use them as features [34]. For extracting those LSP, a numberof points is selected randomly from the 3D model. Then the data points which are closertotheseedpointthanadenedradiusareselected. PCAisthenappliedtondthemostdescriptivefeaturesintheLSP.ThefeatureextractorrepeatsselectingLSPuntilthedesirednumberoffeatureshasbeenfound. BothapproacheswereevaluatedusingimagesfromUND.Therecognitionratereportedfor[82]is93.98%andtherecognitionratereportedfor[34] is93.5%. However, noneoftheapproacheshasbeentestedwithposevariationanddierentscaling.Zhenget al. extracttheshapeindexateachpointinthe3Dmodel anduseitforprojecting the 3D model to 2D space [83]. The 3D shape index at each pixel is represented24byagreyvalueatthecorrespondingpositioninthe2Dimage. ThenSIFTfeaturesareextractedfromtheshapeindexmap. Foreachof theSIFTpointsalocal coordinatesystem is calculated where the z-axis correspondents to the feature points normal. Hencethez-values of theinput imagearenormalizedaccordingtothenormal of theSIFTfeaturepoint theywereassignedto. As soonas thezvalues havebeennormalized,theyaretransformedintoagreylevelimage. Asaresult,Zhengetal. getalocalgreylevel imageforeachoftheselectedSIFTfeatures. NextLBPareextractedforfeaturerepresentation in each of these local grey level images. Comparison is rst performed bycoarsely comparing the shape indexes of key pints and then using Earth movers distanceforcomparingLBPhistogramsfromthecorrespondingnormalizedgreyimages. Zhenget al. evaluatedtheirapproachonasubsetof theUND-J2Collectedandachievedarank-1performanceof96.39%.Figure9:Exampleforlocalsurface(LSP)patchfeaturesasproposedin[34]6 OpenchallengesandfutureapplicationsAs the most recent publications on 2D and 3D ear recognition show, the main applicationof this technique is personal identication in unconstrained environments. This includesapplicationsforsmartsurveillance, suchasin[84] butalsotheforensicidenticationof perpetrators onCCTVimages or for border control systems. Traditionallytheseapplicationeldsarepartoffacerecognitionsystemsbutastheearislocatednexttothe face, it can provide valuable additional information to supplement the facial images.Multimodalearandfacerecognitionsystemscanserveasameansofachievingposeinvarianceandmorerobustness against occlusioninunconstrainedenvironments. Inmost publicvenues surveillancecameras arelocatedoverheadinorder tocaptureasmanypersonsaspossibleandtoprotectthemfromvandalism. Inaddition,mostofthepersonswillnotlookstraightintothecamera,soinmostcasesnofrontalimagesofthepersonswill beavailable. Thisfactposesseriousproblemstobiometricsystems, usingfacial featuresforidentication. If thefaceisnotvisiblefromafrontal angle, theearcanserveasavaluableadditionalcharacteristicinthesescenarios.Because of the physical proximity of the face and the ear, there are also many possibil-itiesforthebiometricfusionofthesetwomodalities. Faceandearimagescanbefusedon the feature level, on the template level and on the score level. Against the background25of thisapplication, therearesomeunsolvedchallenges, whichshouldbeaddressedbyfutureresearchinthiseld.6.1 AutomaticEarLocalizationThe fact that many systems presented in literature use pre-segmented ear images shows,that theautomaticdetectionof ears especiallyinreal-lifeimages is still anunsolvedproblem. Ifearrecognitionsystemsshouldbeimplementedinautomaticidenticationsystems, fast andreliableapproaches for automaticear detectionareof importance.Asarststeptowardsthisgoal,someresearchgroupshavepublisheddatacollections,whichsimulatetypicalvariationsinuncontrolledenvironmentssuchasvaryinglightingconditions, poses and occlusion. Based on these datasets, existing and future approachestoearrecognitionshouldbetestedunderrealisticconditionsinordertoimprovetheirreliability.Moreover, 3Dimagingsystemsbecomeincreasinglycheapinthelastyears. Conse-quently3Dearrecognitionbecomesimportantandwithittheneedoflocatingearsindepth images or 3D models. Currently, only one approach for ear detection in depth hasbeenpublished,whichisarststeptowardseardetectionin3Dimages.6.2 OcclusionandPoseVariationsIn contrast to the face, the ear can be partially or fully covered by hair or by other itemssuch as headdresses, hearing aids,jewelry or headphones. Because of the convex surfaceoftheouterear,partsofitmayalsobeoccludedifthesubjectsposechanges. Insomepublications, robustness against occlusion is explicitly addressed, but there are no studiesontheeectoftheeectofcertaintypesofocclusionlikehairorearringsontherecog-nition rate of an ear recognition system. Once more, the availability of public databaseswhichcontainoccludedearimagesislikelytofosterthedevelopmentof solutionsforposeinvariantandrobustalgorithmsforeardetectionandfeatureextraction.Moreovertoourbestknowledgetherearenostudiesaboutthevisibilityoftheouterear in dierent public environments. In order to develop algorithms for ear detection andrecognition,furtherinformationaboutcommonlyoccludedpartyoftheearisneeded.Occlusion due to pose variations is another unmet challenge in ear recognition system.Similarlytofacerecognition,partsoftheearcanbecomeoccludediftheposechanges.Recently, some feature extraction methods have been proposed, which are robust againstposevariations tosomedegree. However, this issueis not fullysolvedyet. Anotherpossibilityforcompensatingposevariationscouldbetheusageof3Dmodelsinsteadofdepthimagesofphotographs.6.3 ScalabilityCurrentlyavailabledatabases onlyconsist of less than10000ear images. TheonlyexceptionistheUSTBIVcollection,whichhasnotbeenreleasedforthepublicyet. Inrealisticenvironmentsthesizeofthedatabasewillbesignicantlylarger,whichmakesexhaustive search in identication scenarios infeasible. Therefore, not only the accuracy26butalsothecomparisonspeedofearrecognitionsystemswill beinterestingforfutureresearch.In order to make ear recognition applicable for large scale systems, exhaustive searchesshould be replaced by appropriate data structures allowing logarithmic time complexityduringthesearch. Thiscouldforexamplebeachievedbyexploringthepossibilitiesoforganizingeartemplatesinsearchtrees.6.4 UnderstandingSymmetryandAgeingBecauseearrecognitionisoneof thenewereldsof biometricresearch, thesymmetryof the left and the right earhas not beenfullyunderstoodyet. The studies of Iannarelliindicatethatsomecharacteristicsoftheouterearcanbeinheritedandageingslightlyaectstheappearanceoftheouterear. Bothassumptionscouldbeconrmedinmorerecent studies, but because of a lack of sucient data, the eect of inheritance and ageingontheouterearsappearanceisnotfullyunderstoodyet. Furthermore, therearenolargescalestudiesofthesymmetryrelationbetweentheleftandtherightearyet.Thereforeanotherinterestingeldforfutureresearchcouldbe, togainadeeperun-derstanding of the eect of inheritance any symmetry on the distinctiveness of biometrictemplate. Moreover, long term studies on the eect of time on ear templates are neededinordertogetabetterunderstandingofthepermanenceofthischaracteristic.7 SummaryWe have presented a survey on the state of the art in 2D and 3D ar biometrics, coveringeardetectionandearrecognitionsystems. Wecategorizedthelargenumberof2Dearrecognition approaches into holistic, local, hybrid and statistical methods, discussed theircharacteristicsandreportedtheirperformance.Ear recognition is still a new eld of research. Although there is a number of promisingapproaches, noneof themhas beenevaluatedunder realisticscenarios whichincludedisruptive factors like pose variations, occlusionandvaryinglightingconditions. Inrecentapproaches, thesefactorsaretakenunderaccount, butmoreresearchonthisisrequired until ear recognition systems can be used in practice. The availability of suitabletestdatabases,whichwerecollectedunderrealisticscenarios,willfurthercontributetothematurationoftheearasabiometriccharacteristic.Wehavecollectedastructuredsurveyof availabledatabases, existingeardetectionandrecognitionapproachesandunsolvedproblemsforearrecognitioninthecontextofsmartsurveillancesystem, whichweconsidertobethemostimportantapplicationforearbiometrics. Wethinkthatthisnewcharacteristicisavaluableextensionforfacerecognitionsystemsonthewaytoposeinvariantautomaticidentication.27References[1] Abaza A, Ross A.Towards understanding the symmetry of human ears: A biomet-ricperspective. In: FourthIEEEInternationalConferenceonBiometrics: TheoryApplicationsandSystems(BTAS);2010..[2] Bertillon A. La Photographie Judiciaire: Avec Un Appendice Sur La ClassicationEtLIdenticationAnthropometriques. Gauthier-Villars,Paris;1890.[3] Imhofer R. Die Bedeutungder Ohrmuschel f ur die Feststellungder Identitat.Archivf urdieKriminologie.1906;26:150163.[4] IannarelliAV. Earidentication. ParamontPublishingCompany;1989.[5] MeijermanL, Sholl S, Conti FD, GiaconM, vander Lugt C, Drusini A, et al.Exploratorystudyonclassicationandindividualisationof earprints. ForensicScienceInternational.2004;140(1):9199.[6] SinghP, Purkait R. Observations of external earAnIndianstudy. HOMO-Journal of Comparative HumanBiology. 2009;60(5):461472. Available from:http://www.sciencedirect.com/science/article/pii/S0018442X09001164.[7] SforzaC, Grandi G, Binelli M, Tommasi DG, Rosati R, FerrarioVF. Age- andsex-relatedchanges inthe normal humanear. Forensic Science International.2009;187(1-3):110.e1110.e7. Availablefrom: http://www.sciencedirect.com/science/article/pii/S0379073809000966.[8] Meijerman L, Van Der Lugt C, Maat GJR. Cross-Sectional Anthropometric StudyoftheExternalEar. JournalofForensicSciences.2007;52(2):286293.[9] IbrahimMIS,NixonMS,MahmoodiS. Theeectoftimeonearbiometrics. In:InternationalJointConferenceonBiometrics(IJCB);2011.p.16.[10] HoogstrateAJ, Heuvel HVD, HuybenE. Earidenticationbasedonsurveillancecameraimages. Science&Justice.2001;41(3):167172.[11] AlberinkI,RuifrokA. PerformanceoftheFearIDearprintidenticationsystem.ForensicScienceInternational.2007;166(2-3):145154.[12] IslamSMS, BennamounM, Owens R, Davies R. BiometricApproaches of 2D-3DEar andFace: ASurvey. In: SobhT, editor. Advances inComputer andInformationSciencesandEngineering.SpringerNetherlands;2008.p.509514.[13] ChorasM. ImageFeatureExtractionMethodsforEarBiometricsASurvey. In:Computer InformationSystems andIndustrial Management Applications, 2007.CISIM07.6thInternationalConferenceon;2007.p.261265.28[14] PunKH, MoonYS. RecentAdvances in EarBiometrics. In: Automatic Face andGestureRecognition, 2004. Proceedings. SixthIEEEInternational Conferenceon;2004.p.164169.[15] LammiHK. EarBiometics. LappeenrantaUniversityofTechnology,DepartmentofInformationTechnology;2004.[16] RameshKP,RaoKN. Patternextractionmethodsforearbiometrics-Asurvey.In: NatureBiologicallyInspiredComputing, 2009. NaBIC2009. WorldCongresson;2009.p.16571660.[17] Abaza A, Ross A, Herbert C, Harrison MAF, Nixon M. A Survey on Ear Biometrics;2011. AcceptedatACMComputingSurveys. Availablefrom: http://eprints.ecs.soton.ac.uk/22951/.[18] Yan P, Bowyer KW.Biometric Recognition Using 3D Ear Shape.Pattern AnalysisandMachineIntelligence.2007August;29:12971308.[19] FaltemierTC, BowyerKW, FlynnPJ. RotatedProleSignaturesforrobust3Dfeaturedetection. In: 8thIEEEInternationalConferenceonAutomaticFaceandGestureRecognition;2008..[20] Frejlichowski D, TyszkiewiczN. TheWestPomeranianUniversityof TechnologyEarDatabase AToolforTestingBiometricAlgorithms. In: CampilhoA,KamelM, editors. Image Analysis and Recognition. vol. 6112 of Lecture Notes in ComputerScience.SpringerBerlin/Heidelberg;2010.p.227234.[21] Kumar A, WuC. Automatedhumanidenticationusing ear imaging. Pat-ternRecogn.2012March;45(3):956968. Availablefrom: http://dx.doi.org/10.1016/j.patcog.2011.06.005.[22] PrakashS, GuptaP. AnEcient Ear RecognitionTechniqueInvariant toIllu-minationandPose. TelecommunicationSystemsJournal, special issueonSignalProcessingApplicationsinHumanComputerInteraction.2011;30:3850.[23] GrgicM, DelacK, GrgicS. SCfacesurveillancecamerasfacedatabase. Multi-mediaToolsAppl.2011February;51(3):863879. Availablefrom: http://dx.doi.org/10.1007/s11042-009-0417-2.[24] Al Nizami HA, Adkins-Hill JP, Zhang Y, Sullins JR, McCulloughC, CanavanS, etal. Abiometricdatabasewithrotatingheadvideosandhand- drawnfacesketches. In: Proceedingsofthe3rdIEEEinternationalconferenceonBiometrics:Theory, applications andsystems. BTAS09. Piscataway, NJ, USA: IEEEPress;2009.p.3843. Availablefrom: http://dl.acm.org/citation.cfm?id=1736406.1736412.[25] RaposoR, HoyleE, PeixinhoA, ProencaH. UBEAR: Adataset of ear imagescapturedon-the-moveinuncontrolledconditions. In: Computational Intelligence29inBiometricsandIdentityManagement(CIBIM),2011IEEEWorkshopon;2011.p.8490.[26] ChenH, BhanuB. ShapeModel-Based3DEarDetectionfromSideFaceRangeImages. In: IEEEComputerSocietyConferenceonComputerVisionandPatternRecognition-Workshops(CVPR);2005.p.122.[27] Chen H, Bhanu B. Contour Matching for 3DEar Recognition. In: Pro-ceedings of the Seventh IEEEWorkshop on Applications of Computer Vision(WACV/MOTION);2005..[28] Chen H, Bhanu B. Human Ear Recognition in 3D. IEEE Transactions on PatternAnalysisandMachineIntelligence.2007April;29(4):718737.[29] Attarchi S, Faez K, Raei A. A New Segmentation Approach for Ear Recognition.In: Blanc-TalonJ, BourennaneS, PhilipsW, PopescuD, ScheundersP, editors.AdvancedConceptsforIntelligent VisionSystems. vol. 5259of LectureNotesinComputerScience.SpringerBerlin/Heidelberg;2008.p.10301037.[30] AnsariS,GuptaP. LocalizationofEarUsingOuterHelixCurveoftheEar. In:International ConferenceonComputing: TheoryandApplications; 2007. p. 688692.[31] PrakashS, GuptaP. AnEent Ear LocalizationTechnique. ImageandVisionComputing.2012;30:3850.[32] PrakashS, GuptaP. AnEcientTechniqueforEarDetectionin3D: InvarianttoRotationandScale. In: The5thIAPRInternationalConferenceonBiometrics(ICB);2012..[33] AbazaA, HebertC, HarrisonMAF. FastLearningEarDetectionforReal-timeSurveillance. In: FourthIEEEInternational ConferenceonBiometrics: TheoryApplicationsandSystems(BTAS2010);2010.p.16.[34] Islam S, Davies R, Bennamoun M, Mian A. Ecient Detection and Recognition of3DEars. InternationalJournalofComputerVision.2011;95:5273.[35] YuanL, MuZC. EarDetectionBasedonSkin-ColorandContourInformation.In: International Conference on Machine Learning and Cybernetics. vol. 4; 2007. p.22132217.[36] AlvarezL, GonzalezE, MazorraL. Fittingear contour usinganovoidmodel.In: 39thAnnual 2005International CarnahanConferenceonSecurityTechnology(CCST05);2005..[37] ZhouJ,CadavidS,Abdel-MottalebM. HistogramsofCategorizedShapesfor3Deardetection. In: International ConferenceonBiometrics: TheoryApplicationsandSystems;2010..30[38] Arbab-ZavarB,NixonM. OnShape-MediatedEnrolmentinEarBiometrics. In:BebisG, BoyleR, ParvinB, KoracinD, ParagiosN, TanveerSM, etal., editors.AdvancesinVisual Computing. vol. 4842of LectureNotesinComputerScience.SpringerBerlin/Heidelberg;2007.p.549558.[39] Alastair H, Cummings AH, Nixon MS, Carter JN.A Novel Ray Analogy for Enrol-ment of Ear Biometrics. In: Fourth IEEE International Conference on Biometrics:TheoryApplicationsandSystems(BTAS);2010..[40] KumarA, HanmandluM, KuldeepM, GuptaHM. AutomaticEarDetectionforOnline Biometric Applications. In: Third National Conference on Computer Vision,PatternRecognition, ImageProcessingandGraphics(NCVPRIPG); 2011. p. 146149.[41] Hurley DJ, Nixon MS, Carter JN. Force Field Energy Functionals for Image FeatureExtraction. ImageandVisionComputing.2002;20(5-6):311317.[42] Abdel-MottalebM, ZhouJ. HumanEarRecognitionfromFaceProleImages.In: ZhangD, JainA, editors. AdvancesinBiometrics. vol. 3832ofLectureNotesinComputerScience.SpringerBerlin/Heidelberg;2005.p.786792.[43] DongJ, MuZ. Multi-PoseEar RecognitionBasedonForceFieldTransforma-tion. In: SecondInternational SymposiumonIntelligentInformationTechnologyApplication(IITA).vol.3;2008.p.771775.[44] WangX, YuanW. Gabor WaveletsandGeneral Discriminant Analysisfor EarRecogniton. In: 8th World Congress on Intelligent Control and Automation(WCICA);2010.p.6305.[45] Arbab-Zavar B, Nixon M. Robust Log-Gabor Filter for Ear Biometrics. In: Inter-nationalConferenceonPatternRecognition(ICPR);2008..[46] Abate AF, Nappi M, Riccio D, Ricciardi S. Ear Recognition by means of a RotationInvariantDescriptor. In: 18thInternational ConferenceonPatternRecognition,ICPR2006..vol.4;2006.p.437440.[47] SanaPPRAnupam;Gupta. EarBiometrics: ANewApproach. In: AdvancesinPatternRecognition;2007..[48] De MarsicoM, Michele N, RiccioD. HERO: HumanEar RecognitionagainstOcclusions. In: IEEEComputer SocietyConference onComputer VisionandPatternRecognitionWorkshops(CVPRW);2010.p.178.[49] WangXq, XiaHy, WangZl. TheResearchof EarIdenticationBasedOnIm-provedAlgorithmof MomentInvariants. In: ThirdInternational ConferenceonInformationandComputing(ICIC);2010.p.58.31[50] WangX, YuanW. HumanEar recognitionBasedonBlockSegmentation. In:International Conference on Cyber-Enabled Distributed Computing and KnowledgeDiscovery(CyberC);2009.p.262266.[51] LoweGD. Objectrecognitionfromlocal scale-invariantfeatures. In: IEEEIn-ternational ConferenceonComputerVision(ICCV1999). vol. 2. IEEEComputerSociety;1999.p.11501157.[52] Bustard JD, Nixon MS. Toward Unconstrained Ear Recognition FromTwo-Dimensional Images. Systems, ManandCybernetics, PartA: SystemsandHu-mans.2010April;40:486.[53] Arbab-ZavarB,NixonMS,HurleyDJ. OnModel-BasedAnalysisofEarBiomet-rics.In: First IEEE International Conference on Biometrics: Theory, Applications,andSystems,2007.(BTAS2007);2007.p.15.[54] BadrinathG, GuptaP. FeatureLevel FusedEarBiometricSystem. In: SeventhInternational ConferenceonAdvancesinPatternRecognition(ICAPR); 2009. p.197200.[55] Kisku DR, Mehrotra H, Gupta P, Sing JK.SIFT-based ear recognition by fusion ofdetected keypoints from color similarity slice regions. In: International Conferenceon Advances in Computational Tools for Engineering Applications (ACTEA); 2009.p.380385.[56] BayH, TuytelaarsT, Gool LV. SURF: SpeededUpRobustFeatures. In: Pro-ceedingsofthe9thEuropeanConferenceonComputerVision;2006..[57] ChorasM. PerspectiveMethodsofHumanIdentication: EarBiometrics. Opto-ElectronicsReview.2008;16:8596.[58] Mu Z, Yuan L, Xu Z, Xi D, Qi S. Shape and Structural Feature Based Ear Recog-nition. In: Li S, Lai J, TanT, FengG, WangY, editors. AdvancesinBiometricPersonAuthentication. vol. 3338of LectureNotesinComputerScience. SpringerBerlin/Heidelberg;2005.p.311364.[59] RahmanM, IslamR, BhuiyanNI, AhmedB, IslamA. PersonIdenticationUs-ingEar Biometrics. International Journal of The Computer, The Internet andManagement.2007August;15:18.[60] GuoY,XuZ. EarRecognitionUsingaNewLocalMatchingApproach. In: 15thIEEEInternationalConferenceonImageProcessing(ICIP);2008.p.289292.[61] Wang Zq, Yan Xd. Multi-Scale Feature Extraction Algorithm of Ear Image. In: In-ternational Conference on Electric Information and Control Engineering (ICEICE);2011.p.528.[62] JegesE, MtL. Model-BasedHumanEarLocalizationandFeatureExtraction.IC-MED.2007;1(2):101112.32[63] LiuH,YanJ. Multi-viewEarShapeFeatureExtractionandReconstruction. In:ThirdInternational IEEEConferenceonSignal-ImageTechnologiesandInternet-BasedSystem(SITIS);2007.p.652658.[64] MukundanR, OngSH, LeePA. ImageanalysisbyTchebichef moments. IEEETransactionsonImageProcessing.2001September;10(9):13571364.[65] Lu L, Xiaoxun Z, Youdong Z, Yunde J.Ear Recognition Based on Statistical ShapeModel. In: FirstInternationalConferenceonInnovativeComputing, InformationandControl(ICICIC);2006.p.353356.[66] Yuan L, Mu Z.Ear Recognition Based on 2D Images.In: First IEEE InternationalConferenceonBiometrics: Theory, Applications, andSystems(BTAS); 2007. p. 15.[67] VictorB,BowyerK,SarkarS. Anevaluationoffaceandearbiometrics. In: 16thInternational Conference on Pattern Recognition (ICPR). vol. 1; 2002. p. 429 432vol.1.[68] ChangK,BowyerKW,SarkarS,VictorB. ComparisonandCombinationofEarandFaceImagesinAppearance-BasedBiometrics. IEEETransactionsinPatternAnallysisandMachineIntelligene.2003September;25:11601165.[69] AlarajM,HouJ,FukamiT. Aneuralnetworkbasedhumanidenticationframe-workusingearimages. In: TENCON2010- 2010IEEERegion10Conference;2010..[70] ZhangH,MuZ. CompoundStructureClassierSystemforEarRecognition. In:IEEE International Conference on Automation and Logistics (ICAL); 2008. p. 23062309.[71] Xie Z, MuZ. Ear RecognitionUsing LLEandIDLLEAlgorithm. In: 19thInternationalConferenceonPatternRecognition(ICPR);2008.p.14.[72] Nanni L, Lumini A. A Multi-Matcher For Ear Authentication. Pattern RecognitionLetters.2007December;28:22192226.[73] YuizonoT,WangY,SatohK,NakayamaS. StudyonIndividualRecognitionforEar Images by using Genetic Local Search.In: Proceedings of the 2002 Congress onProcessing Scociety of Japan (IPSJ) Kyushu Chapter Symposium; 2002. p. 237242.[74] Yaqubi M, Faez K, Motamed S.Ear Recognition Using Features Inspired by VisualCortexandSupportVectorMachineTechnique. In: International ConferenceonComputerandCommunicationEngineering(ICCCE);2008.p.533537.[75] GutierrezL, MelinP, LopezM. ModularNeural NetworkIntegratorforHumanRecognitionFromEar Images. In: The2010International Joint ConferenceonNeuralNetworks(IJCNN);2010..33[76] NaseemI,TogneriR,BennamounM. SparseRepresentationforEarBiometrics.In: Bebis G, Boyle R, Parvin B, Koracin D, Remagnino P, Porikli F, et al., editors.AdvancesinVisual Computing. vol. 5359of LectureNotesinComputerScience.SpringerBerlin/Heidelberg;2008.p.336345.[77] YanP,BowyerKW. AFastAlgorithmforICP-Based3DShapeBiometrics. In:FourthIEEEWorkshoponAutomaticIdenticationAdvancedTechnologies;2005.p.213218.[78] CadavidS, Mahoor MH, Abdel-MottalebM. Multi-Modal Biometric ModelingandRecognitionof theHumanFaceandEar. In: IEEEInternational WorkshoponSafety,SecurityRescueRobotics(SSRR);2009.p.16.[79] ZhouJ, CadavidS, Abdel-MottalebM. AcomputationallyEcient Approachto3DEarRecognitionEmployingLocal andHolisticFeatures. In: IEEECom-puter Society Conference on Computer Vision and Pattern Recognition Workshops(CVPRW);2011.p.98105.[80] Passalis G, Kakadiaris IA, Theoharis T, Toderici G, Papaioannou T. Towards Fast3DEarRecognitionForReal-LifeBiometricApplications. In: IEEEConferenceonAdvancedVideoandSignalBasedSurveillance(AVSS2007);2007.p.3944.[81] LiuH, ZhangD. Fast3DPointCloudEarIdenticationbySliceCurveMatch-ing. In: 3rdInternational ConferenceonComputer ResearchandDevelopment(ICCRD);2011.p.224.[82] IslamSMS,BennamounM,MianAS,DaviesR. AFullyAutomaticApproachforHumanRecognitionfromProleImages Using2Dand3DEar Data. In: Pro-ceedingsof3DPVT-theFourthInternatinoalSymposiumon3DDataProcessing,VisualizationandTransmission;2008..[83] Zeng H, Dong JY, Mu ZC, Guo Y. Ear Recognition Based on 3D Keypoint Match-ing. In: IEEE10thInternational ConferenceonSignal Processing(ICSP); 2010.p.1694.[84] CadavidS, Abdel-MottalebM. 3Dearmodelingandrecognitionfromvideose-quencesusingshapefromshading. In: 19thInternational ConferenceonPatternRecognition(ICPR);2008.p.14.[85] IslamSMS, BennamounM, DaviesR. FastandFullyAutomaticEarDetectionUsingCascadedAdaBoost. In: Applicationsof ComputerVision, 2008. WACV2008.IEEEWorkshopon;2008.p.16.[86] Liu H, Liu D.Improving Adaboost Ear Detection with Skin-color model and Multi-templateMatching. In: 3rdIEEEInternationalConferenceonComputerScienceandInformationTechnology(ICCSIT).vol.8;2010.p.106109.34[87] ShihHC, HoCC, ChangHT, WuCS. EarDetectionBasedonArc-MaskingEx-tractionandAdaBoostPollingVerication. In: FifthInternationalConferenceonIntelligentInformationHidingandMultimediaSignal Processing(IIH-MSP). vol.eardetection;2009.p.669672.[88] Burge M, Burger W. 13. In: Jain AK, Bolle R, Pankanti S, editors. Ear Biometrics.SpringerUS;1998.p.273285.[89] MorenoABandSanchez,VelezJF. OnTheUseofOuterEarImagesforPersonalIdenticationinSecurityApplications. In: IEEE33rdAnnual 1999InternationalCarnahanConferenceonSecurityTechnology;2002.p.469476.[90] YuanL, chunMuZ, ZhangY, LiuK. Ear RecognitionusingImprovedNon-Negative Matrix Factorization. In: 18th International Conference on PatternRecognition(ICPR).vol.4;2006.p.501504.[91] WangY,chunMuZ,ZengH. Block-BasedandMulti-ResolitionmethodsforEarrecognition Using Walelste Transform and Uniform Local Binary Patterns. In: 19thInternationalConferenceonPatternRecognition(ICPR);2008.p.14.[92] Fooprateepsiri R, Kurutach W.Ear Based Personal Identication Approach Foren-sicScienceTasks. ChiangMaiJournalofScience.2011January;38(2):166175.[93] IslamSM, DaviesR, MianMASanBennamoun. AFastandFullyAutomaticEarRecognitionApproachBasedon3DLocal SurfaceFeatures. In: Proceedingsof the10thInternational ConferenceonAdvancedConceptsforIntelligentVisionSystems.ACIVS08.Berlin,Heidelberg: Springer-Verlag;2008.p.10811092.35