22-09-2012, 12:56 PM
Reliable detectionofeyefeaturesandeyesincolorfacialimagesusing
ternary eye-verifier
1Reliable.pdf (Size: 2.37 MB / Downloads: 17)
ABSTRACT
Eye detectionplaysanimportantroleinapplicationsrelatedtofacerecognition.Thepositionofeyes
can beusedasareliablereferenceforotherfacialfeaturedetection.Thispaperpresentsanovel
approachforthepreciseandreliabledetectionofeyesbyintroducingaternaryeye-verifier.Initially,
the faceregionisdetectedbycombiningcolorinformationandtheHaar-likefeaturedetector.Theface
region isthenbinarizedandfilteredwithcircularfilterstodetecteyecandidatesatthepeaksinthe
filteredresponse.Eacheyecandidateisfedintoaternaryeye-verifierthatincludesaproposedeye
featureextractorbasedonK-meansclusteringwithcompensationforvarietyiniriscolor.Theeye
templateintheeye-verifierisconstructedbasedonboththeknowledgeofeyegeometryandthe
detectedeyefeatures.ThetemplatematchingismadebytheternaryHammingdistance.Experiments
over acollectionofFERETfacedatabaseandhouse-madefacedatabasewithdifferentheadposes
confirmthattheproposedmethodachievespreciseandreliabledetectionofeyesfromcolorfacial
images withvariationinillumination,pose,eyegazingdirection,andrace.
Introduction
Previousworksoneyedetection
Biometrics isundoubtedlycontributingtothehumanidentity
management inthemodernsociety.Withseveraladvantages
compared tootherbiometrics,facerecognitionisfindingmany
applications, sinceitsbirthabout2decadesago.Robustand
accurate detectionofeyesiscrucialinfacerecognition,asthe
eyes areconsideredasthemostsalientanddistinguishable
feature ofthehumanface [1,2]. Oncethetwoeyesarelocated
reliably, theotherfacialfeatures(e.g.mouth,nose,ears,etc.)can
be roughlyestimatedbasedoneyepositions,astheyhave
reference valuestoeachother [3,4]. Furthermore,locationofeyes
can beusedtodeterminetheposeoftheheadthankstothe
anatomical structureofthehumanface.
Proposedmethodfordetectingeyesandeyefeatures
This paperproposesatemplate-featurebasedeye-verifierto
detect eyesinfacialcolorimages.Theproposedmethodishighly
appropriate forthenewgenerationofhighqualityfacialphotos
used forbiometricpassportsreleasedbytheInternationalCivil
Aviation OrganizationICAOReport [29]. Theproposedmethod
consists oftwostages:eyeregioncandidatedetectionandeye
verification. Thefaceregionisfirstlydetectedbyahybridface
detector whichcombinestheskincolorinYCbCr color spaceand
the Haar-likefeatures [30]. Thereafter,onlythefaceregionis
cropped andthenbinarizedforsubsequentprocesses.Thecircular
filer [31,32] isappliedtothebinaryfaceimage.Aneye-weight
matrix generatedfromthebinaryfaceimage,whichputssmall
weights onthenear-boundaryregionsandeyebrows,aswellas
beards, ifpresent,isweightedtothefilteredimagetoproducean
eye-peaked image.Theeyecandidatesareselectedfromthe
peaks inthefilteredimage.
Previousworksoneyefeaturedetection
Vezhnevets andDegtiarevaproposedin [22,23] amethodto
detect eyefeaturesthatfindstheirisintheredchannelby
searching forthehighlightreflections,andthenfindstheupper
eyelid bylookingforluminancevalleypoints.Finally,thelower
eyelid isestimatedfromtheknownirisandeyecorners.In [23],
Xu etal.proposedtodefinethecanthus(eyecorner)astheangle
formed bytheeyelids.Inthestudy,ananglemodelisfirst
constructed bytwoeyelids,thentwofeatures,onetocharacterize
the appearancedifferencebetweeninnerandoutercanthus
regions, theothertoemphasizetheroleofthecanthusangle
bisector region,arefusedbylogisticregressiontodetectthe
canthus. Zhengetal. [24,25] usetheHchannelofHSVcolorspace
to detecttheiris,andthentheeyecornersarelocatedbyaneye-
corner filtergeneratedfromGaborfeaturespace.Theeyelidsare
then fittedbysplinefunctions.KhosraviandSafabakhsh [26]
detect theirisusingatemplate-matchingschemeinthebinary
image anddetectcornersbyawindowshiftingmethodology.The
sclera oftheeyeisthenlocatedbyaTASOM-ACMalgorithm [27].
In [28].
Testofeye-verifier
Thissectionshowstheperformanceoftheternaryeye-verifier,
assuming theinputisacoloreyecandidate thatneedstoverifyifit
is aneye,astheternaryeye-verifierisastandalonemodule. Fig. 12
showstheproceduretoapplytheeye-verifierfortheeyedetection.
The eyedetectionprocessterminateswheneitherbotheyes
are foundorfoureyecandidateshavepassedthroughtheeye-
verifier, asindicatedin Fig. 12. Thedecisionoftheeye-verifieris
of coursedependentonthethresholdvalueoftheternary
Hamming distance(HD);thereforethenumberofverifications
and averagenumberofverificationsperimagealsodependonthe
threshold valueoftheternaryHammingdistance. Table 2 shows
those numbersforthetestontheFERETcollection.Asthe
Hamming distanceincreases,itislooserfortheeye-verifierto
accept aneye,thusthenumberofverifications(andtheaverage
number ofverificationsperimage)deceases.