Stratified SIFT Matching for Human Iris Recognition

0.41 MB·English

by  Sambit Bakshi

Stratified SIFT Matching for Human Iris Recognition
SambitBakshi,HunnyMehrotra,andBanshidharMajhi
DepartmentofComputerScienceandEngineering,
NationalInstituteofTechnologyRourkela
Odisha,India
[email protected],[email protected],[email protected]
http://www.nitrkl.ac.in
3 Abstract. This paper proposes an efficient three fold stratifiedSIFT matching
1 foririsrecognition. Theobjectiveistofilterwronglypairedconventional SIFT
0 matches.InStrataI,thekeypointsfromgalleryandprobeirisimagesarepaired
2
usingtraditionalSIFTapproach.Duetohighimagesimilarityatdifferentregions

n ofiristheremaybesomeimpairments.Thesearedetectedandfilteredbyfind-
a inggradientofpairedkeypointsinStrataII.Further,thescalingfactorofpaired
J keypoints is used to remove impairments in StrataIII. The pairs retained after

6 Strata III are likely to be potential matches for iris recognition. The proposed
systemperformswithanaccuracyof96.08%and97.15%onpubliclyavailable

] CASIAV3andBATHdatabasesrespectively.Thismarkssignificantimprovement
V
ofaccuracyandFARovertheexistingSIFTmatchingforiris.
C
. Keywords: IrisRecognition,StratifiedSIFT,Keypoint,Matching.
s
c
[
1 Introduction
1
v
Irisisthesphincterhavinguniquefloweryrandompatternaroundthepupil.Itisanin-
8
9 ternalorganwithcomplexuniquefeaturesthatarestablethroughoutthelifetimeofan
9 individual.Therehasbeensignificantresearchdoneintheareaofirisrecognitionusing
0 globalfeatures[1,2,3,4].However,theseapproachesfailtopossessinvariancetoaffine
.
1 transformations,occlusionandrobustnessto unconstrainedirisimages.Thus,there is
0 astringentrequirementtodevelopirisrecognitionsystemsuitablefornon-cooperative
3 images.Keypointdescriptorsareinvarianttoaffinetransformationaswellaspartialoc-
1
clusion.ScaleInvariantFeatureTransform(SIFT)isawellknownkeypointdescriptor
:
v for object recognition [5]. Due to inherent advantages, SIFT is capable of perform-
i
X ingrecognitionusingnon-cooperativeirisimages[6].InSIFTmatchingapproach,the
differenceof Gaussianimagesareused toidentifykeypointsatvaryingscale andori-
r
a entation.Theorientationis assignedto eachdetectedkeypointanda windowistaken
relativetodirectionoforientationtofindthedescriptorvector.Duringrecognition,the
keypointsaredetectedfromgalleryandprobeimagesandmatchingisperformedusing
nearestneighbourapproach.ThechallengewithconventionalSIFTmatchingwhenap-
plied to iris recognitionis to find texturesimilarity between same regionsof two iris.
SIFT fails as itdoesnotconsiderspatialinformationofthe keypoints.To makeSIFT
functionalforiris,anovelmatchingapproachhasbeendevelopedthatcombinesspatial
informationalongwithlocaldescriptorofeachkeypoint.
(a) StrataI:η=98 (b) StrataII:η=65
(c) StrataIII:η=54
Fig.1.Matches(η)atdifferent strataobtained betweentwoinstancesof sameindividual taken
fromCASIAV3
Theorganizationofthepaperisasfollows.ConventionalSIFTmatchingapproach
forlocaldescriptorsisdiscussedinSection2.ProposedstratifiedSIFTmatchingisex-
plainedindetailinSection3.Thejoboffilteringisdoneintwosequencialsteps:Gra-
dientbased filtering and Scale based filtering. Section 4 providesthe results obtained
usingproposedapproach.Finally,conclusionsaregivenattheendofthepaper.
2 Conventional SIFTMatching
Thematchingalgorithmplaysasignificantroleinanybiometricsystemasitactslike
aonewaygatewaythroughwhichonlygenuinematches(iftwoimagesarefromsame
subject) will pass and imposter matches (if two images are from different subjects)
are blocked. In local feature matching, the total number of paired keypoints is used
to find the authenticity of an individual. Let I be the set of all images available in
the iris database. For understanding, I be a gallery iris image and I be a probe iris
m n
imagewhere I ,I ǫ I. The matchingalgorithmvalidates I against I . Inthe conven-
m n n m
tional SIFT matching, for each keypoint in I the Euclidean distance is found with
m
everykeypointin I . Thenearestneighbourapproachpairstheith keypointin I with
n m
jth keypointin I ,iffthedescriptordistancebetweenthetwo(aftermultiplyingwitha
n
threshold)isminimum[5].Thetwokeypointsarepairedandremovedfromthesetof
keypointsdetectedfromI andI .Thisprocessisiteratedfortheremainingkeypoints
m n
untilanytwo keypointscan be matched.Thisapproachperformsmoderatelywellfor
unconstrainedirisrecognition[6].However,asSIFTdeterminesimagesimilarityusing
128-dimensionallocalfeaturesonly,henceitmaywronglypair(impairment)somekey-
pointsforiris.Thus,theexistingapproachismodifiedusingtwostratawhichremoves
impairedmatchesusingspatialinformationofthematchingkeypointsandcontributes
inachieveingbetterrecognitionaccuracy.
3 Stratified SIFTMatching
IntheproposedpaperaneffortismadetoimprovisetheconventionalSIFTmatching.
Thepupilandiriscirclesareassumedtobeconcentric,hencealllocalizedimageshave
pixelsize 2r×2r, where r is the radiusofiris. Thepupilcenter aswell as iriscentre
are located at (r,r). Therefore the localized images do not have transformation due
to translation. However,there is a possibility of iris images being transformeddue to
rotation(tiltofsubject’shead),scaling(changeincameratoeyedistance)orboth[6].
TheSIFTmatchingalgorithmmatcheskeypointsthathavesimilaritybetweenthelocal
descriptors(asdiscussedinSection2)butfailstoconformtospatialrelationship.The
removal(filtering)ofimpairmentsbytheproposedapproachretainsonlythosematches
thataremoreprobabletobepotentialmatches.LetK bethesetofm keypointsfound
m 1
in I and K be the set of n keypointsfoundin I by applyingSIFT detector. These
m n 1 n
setsofkeypointsareusedtocomprehendthestratifiedSIFTmatching.
3.1 StrataI:SIFTMatching
Let R be the orderedset containingthe matches between K and K by conventional
m n
SIFTmatchingasdiscussedinSection2.Hence,Rcontainsonlythosepairs(i, j)where
ithkeypointinK ismatchedwith jthkeypointinK asshowninFig.1(a).Letηbethe
m n
numberofmatchesfoundwhereηǫ[0,min(m ,n )].AssetRisgeneratedsolelyonthe
1 1
basisoflocaldescriptorpropertyitmaywronglypairkeypointsfromdifferentregions
of iris. Hence, there is a need to combinespatial informationwith local descriptorto
filteroutimpairedkeypointsasdiscussedinsubsequentstrata.
3.2 StrataII:GradientBasedFiltering
In this strata gradientbased filtering is performedto removeimpairmentsfromR. To
compute gradient for each pair of keypoints (i, j) in R, the angles are obtained from
respectiveimagecenters(r,r).Thus,θ iscomputedfromI andφ iscomputedfrom
i m j
I .Theangleofrotationforkthpairiscalculatedasγ =(φ −Θ)mod360°(depictedin
n k j i
Fig.3(a)).ConsideringSIFTtobecompletelyflawless(duetorobustnessproperty,no
falsematchisfound)andefficient(duetopropertyofrepeatability,allpossiblematches
arefound)[5];thevalueofγ derivedshouldbesame∀k.Butinpractice,SIFTdoesnot
k
givesuchprecisematches.Thus,itisdifficulttoobtainuniquevalueofγevenwhenI
m
andI belongtothesamesubject.Ratheradistributionofγisobtained.Ahistogramis
n
plottedwithx-axiscomprisingbinswitharangeofvaluesofγ,andy-axiscomprising
numberofmatchesfallinginaparticularbinasshowninFig.2.Thenumberofbinsin
thehistogram(nobins)issubjecttoimplementationissue.Inproposedsystem,nobins
istakenas10.Thedistributionofγgivesasinglepeakincasethetwoirisimages(I
m
andI )arefromthesamesubject.Incontrast,nodistinctpeakshouldbefoundincase
n
thetwoirisimagesarefromdifferentsubjects.Theremaybeerrorduetodiscretization
ofbins,sotwoadjacentbinsofthepeakarecombinedtoimprovepeakdensity(number
ofmatches).Theideaistofindwhetherthedensityofthepeakexceedstheboundary
criteria.Itisinferredthata peakisstrongif thedensityexceedscertainhigherbound
(hp%oftotalnumberofmatches).
40 Peak
35
30
es
atch25 Rejected Matches
M
of 20
er
b
m15
u
N
10
5
0
0 36 72 108 144 180 216 252 288 324 360
γ (in degree)
Fig.2.Distributionofγfornumberofmatchesbetweentwoinstancesofsameindividualfrom
CASIAV3
Likewise peak is weak if the density is less than lower bound(lp% of total num-
ber of matches). If a strong peak is found, an angular range is specified around the
peak.ThosematchesinRforwhichγarenotwithintheangularrangearefoundtobe
impairedandremovedfromRtogenerateRinter.
θ
i φ
j
d
1 d
2
(a) γ=(φ −θ)mod360° (b) ψ= d2
j i d1
Fig.3.(a)GradientcomputationinStrataII,(b)LocalscalingfactorcomputationinStrataIII
For example, as shown in histogram in Fig. 2, the peak is found at 0th bin which
represents gradient value of 0° to 36° with a central value of 18°. Hence only those
pairshavingangularrangebetween(18±90)mod360°areretained.Thus,itisevident
thatRinter ⊆ Rafterremovingsomeimpairments.Fig.1(b)showspairedkeypointsin
Rinterwithconsiderablereductionofη.Ifnostrongpeakisfounditisinferredthatall
matchesinRarefaulty,andremoved.AsaresultRinterbecomesempty.
3.3 StrataIII:ScaleBasedFiltering
In this strata further filtering of Rinter is performed on the basis of global and local
scaling factor between the gallery and probe images. The global scaling factor (sf)
betweentwoimagesisdefinedasratioofprobeirisradius(r )togalleryirisradius(r ).
n m
A scale rangewithcertaintolerancearound sf isempiricallytakento handlealiasing
artifact.Fromimplementationperspective,thescalerangeistakenas(sf±0.2).Inideal
case,ifgalleryandprobebelongtosameindividualthescalingfactorbetweenallpaired
keypointsshouldbeunique.However,thisdoesnotholdgoodinpracticalscenarios.
Duringfiltering,foreachelementinRinter,twoEuclideandistancesarecalculated-
(a)d :distanceofith keypointofI fromitscenterand(b)d :distanceof jth keypoint
1 m 2
of I fromits center. Localscaling factor (ψ) foreach elementof Rinter is calculated
n
as ψ = d /d (shown in Fig. 3(b)). Matches having ψ within scale range discussed
2 1
abovequalifiestobepotentialandstoredinRnew,elsearelabeledasfaultyandfiltered.
Fig.1(c)showspairedkeypointsinRnewafterfurtherreductionofη.
4 EXPERIMENTAL RESULTS
The proposed stratified SIFT matching is tested on publicly available BATH [8] and
CASIAV3 (CASIAV3) [9] databases. Database available from BATH University in-
cludesimagesfrom50subjects(20imagespersubjectfromboththeeyes).CASIAV3
(CASIAV3)comprises249subjectswithtotalof2655imagesfromboththeeyes.The
experiments are carried out on 2.13GHz Intel(R) CPU using Matlab. To validate the
system performancesome standard error measures [7] are used1. The results are car-
riedoutinthreedifferentstrataasgiveninTable1.InStrataI,thetwoirisimagesare
matchedusingconventionalSIFTapproach.Thisapproachperformswithanaccuracy
of85.81%onCASIAV3database.Likewise,forBATHdatabaseanaccuracyof97.04%
Table1.PerformancemeasuresforstratifiedSIFTmatching
Databases→ CASIAV3 BATH
Approach↓ FAR FRR ACC d’ FARFRR ACC d’
StrataI(ConventionalSIFT) 17.4810.9185.811.20 1.57 4.35 97.042.73
StrataII(GradientbasedFiltering) 2.49 5.45 96.032.46 0.97 6.09 96.472.81
StrataIII(ScalebasedFiltering) 2.39 5.45 96.082.20 1.34 4.35 97.152.90
is obtained.To improvethe performanceof the system, the objectiveof the proposed
researchis to reducefalse acceptances.In Strata II, the impairmentsare removedus-
ing gradient filtering which significantly increases the seperability measure between
false andgenuinematchesas indicatedbyd’ valuesgivenin the Table 1. Furtherim-
provementinseparabilityandaccuracyarebroughtbyscalefilteringinStrataIII.The
accuracyvaluesareplottedagainstchangeinnumberofmatchesasshowninFig.4(a).
1FAR:FalseAcceptanceRate,FRR:FalseRejectionRate,ACC:Accuracy,d’:d-primevalue
100 100
Strata I Strata I
95 Strata II 90 Strata II
Strata III Strata III
90 R) 80
Accuracy (in %) 667788050505 False Rejection Rate (FR 234567000000
55 10
50 0
0 50 100 150 200 10−2 10−1 100 101 102
Matching score False Acceptance Rate (FAR)
(a) (b)
Fig.4.(a)Accuracycurveforthreestrata,(b)ROCcurveforthreestrataonCASIAV3
The Receiver Operating Characteristic (ROC) curves [7] for three different strata
are shown in Fig. 4(b). The distribution of genuine and imposter scores after Strata
III is shown in Fig. 5. All graphical results are obtained for CASIAV3 and similar
observationsaremadeforBATHdatabase.

Imposters
Genuine
103
n
o
uti
strib102
Di
e
or
c
S
101
100
0 20 40 60 80 100 120 140 160
Matching Score
Fig.5.FinalhistogramofscoresonCASIAV3
5 CONCLUSIONS
In this paper, a novel stratified SIFT matching technique is proposed that improvises
conventional SIFT by removing wrong pairs. This approach provides boost in accu-
racyduetoconsiderablereductioninFAR.TheFARisreducedby15.09%and0.23%
for CASIAV3 and BATH respectively. From the results it has been observed that the
proposedalgorithmiscompletelyflawless,i.e.,matchesremovedareguaranteedtobe
wrongmatcheswhereasitisnotcompletelyefficient,i.e.,allimpairmentsbySIFTare
notguaranteedtobefiltered.However,thegaininaccuracyissubstantialwhichmarks
itsapplicabilityforunconstrainedirisrecognition.
References
1. Daugman,J.:Theimportanceofbeingrandom:Statisticalprinciplesofirisrecognition.Pat-
ternRecognition36(2),pp.279291(2003)
2. Kim,J.,Cho,S.,Choi,J.,Marks,R.J.:IrisRecognitionUsingWaveletFeatures.The.Journal
ofVLSISignalProcessing(38),147156(2004)
3. Monro,D.M.,Rakshit,S.,Zhang,D.:Dct-basedirisrecognition.IEEETransactionsonPattern
AnalysisandMachineIntelligence29(4),586595(2007)
4. Tisse,C.,Martin,L.,Torres,L.,Robert,M.:Personidentificationtechniqueusinghumaniris
recognition.In:Proc.ofVisionInterface,pp.294299(2002)
5. Lowe,DG:Distinctiveimagefeaturesfromscale-invariantkeypoints.InternationalJournalof
ComputerVision6(2),91110(2004)
6. Belcher, C., Du, Y.: Region-based sift approach to iris recognition. Optics and. Lasers in
Engineering47(1),139147(2009)
7. Jain,A.,Flynn,P.,Ross,A.:HandbookofBiometrics.Springer-VerlagNewYork,.Inc.,Se-
caucus,NJ,USA(2007)
8. BathUniversityDatabase,http://www.bath.ac.uk/elec-eng/research/sipg/irisweb
9. ChineseAcademyofSciences’InstituteofAutomation(CASIA)IrisImageDatabaseV3.0,
http://www.cbsr.ia.ac.cn/english/IrisDatabase.asp