人脸表情识别文献_「金点文库网」分享知识创造价值

赞颂老师的名言-一年级教学反思

JournalofDonghuaUniversity（Eng．Ed．）Vol．29，N o．1（2012）
71
FacialExpressionRecognitionBase donthe
Q-
shiftDT-
CWT
andRotation
InvariantLBP
CHENLei（
陈蕾
）
*
School ofElectronics＆InformationEngineering，SoochowUniver sity，Suzhou215021，China
Abstract：Inthispaper，an ovelmethodbasedondual- treecomplex
wavelettransform（DT-CWT）androtation invariantlocalbinary
pattern（LBP）forfacialexpre ssionrecognitionisproposed．The
quartersampleshi ft（Q-shift）DT-CWTcanprovideagroupdelay
of14ofas ampleperiod，andsatisfytheusual2-bandfilterbank
constraintsofnoaliasingandperfectreconstruction．To resolve
illuminationvariationinexpressionverifi cation，low-frequency
coefficientsproducedbyDT- CWTaresetzeroes，high-frequency
coefficientsareu sedforreconstructingtheimage，andbasicLBP
histog ramismappedonthereconstructedimagebymeansof
his togramspecification．LBPiscapableofencodingtexturea nd
shapeinformationofthepreprocessedimages．Theh istogramgraphs
builtfrommulti-scalerotationinva riantLBPsarecombinedtoserve
asfeatureforfurther recognition．Templatematchingisadoptedto
classif yfacialexpressionsforitssimplicity．Theexperimental results
showthattheproposedapproachhasgoodperfo rmanceinefficiency
andaccuracy．
Keywords：fac ialexpressionrecognition；dual- treecomplexwavelet
transform（DT-CWT）；localbinar ypattern（LBP）；histogram；
similaritymeasure
C LCnumber：TP391Documentcode：A
ArticleID：1672-522 0（2012）
01-0071-05
，WANGJia-jun（
王加俊
），SUNBing（
孙兵
）
Introduction
Facialexpr essionrecognitionistoanalyzeanddetectthe
specia lexpressionstatefromgivenexpressionimagesorvideoframesandthentoascertainthesubject'sspecificinbo rn
emotion，achievingsmarterandmorenaturalintera ction
betweenhumanbeingsandcomputers．Inhumantoh uman
interaction，Mehrabiandiscoveredthatverbalc uesprovided7%
ofthemeaningofthemessage；vocalcue s，38%；andfacial
expressions，55%
［1］
．Thus facialexpressionprovidesmore
informationaboutth einteractionthanthespokenwords．
Automaticfacial expressionrecognitionplaysanimportantrole
inthe developmentofpatternrecognition，computervision，
computergraphics，artificialintelligence，physiolog y，
psychologyandsoon．Thestudyoffacialexpression
recognitionhasfounditsvaluesineconomyandsociet y．
Duetoitsapplicationsonsociologyandcomputervi sion，
automaticfacialexpressionrecognitionhasat tractedmoreand
moreattention．In1978，onthebasiso ftheanatomy，Ekman
andFriesenbuiltthefaceactionc odingsystem（FACS）that
associatedfacialexpressio nwithmusclemovement
［2］
．By
FACS，encoding allpossiblefaceexpressionsbecameareality．
Theni n1984，EkmanandFriesenputforwardthatthe
combinat ionofspecificFACSactionunitscouldindicatefacial
expressionofemotions
［3］
．Accordingtodiffer entemotions，
facialexpressionscanbedividedintos ixtypes：happy，sad，
surprise，fear，anger，anddisgu st
［4，5］
．Thesixbasictypesare
agreedwidel ybyresearchersandtreatedasthefacialexpression
c ategories．Facialexpressionrecognitiongenerallyincl udes
threestages：facedetectionandlocalization，t hefacialfeature
extraction，andexpressionrecogni tion．Facialexpressionisvery
complex，forexample， iftheopeningmouthdoesnotrepresent
smilenecessar ily，itmaybecryorsurprise，andthesamekind
ofexpre ssionsmayhavealotofdifferentways，suchashappy，
s omeareopenmouthcachinnation，someareclosedmouth
smile．Therefore，thefacialexpressionanalysisisadiff icult
taskthatmainlyreflectsintheaccuracyexpres sionfeature
extractionandvalidityofexpressionfe atureextraction．
Researchershavemadesomeachieve mentsinfacialexpression
analysis．Therearethefol lowingmethods：principalcomponent
analysis（PCA）， independentcomponentanalysis（ICA），linear
discri minantanalysis（LDA），fisherlinearjudgingmethod，
clusteringdiscriminantanalysis，elasticchartmatchin gmethod，
gaborwaveletmethod，andlocalprincipalco mponentanalysis，
etc．Everyalgorithmmentionedhas someeffects，butisnot
verysatisfactoryandneedsto beimproved．
Inthispaper，abriefreviewofDT-CWTand localbinary
patterns（LBPs）isintroducedandanovel methodforfacial
expressionfeatureextractionispr oposed．Bythismethod，the
approximationcoefficien tsderivedfromDT-CWTareresetand
detailscoefficie ntsarepreserved．InverseofDT-DWTis
employedbyusi ngmodifiedcoefficients，andthereconstructed
imag eisreferredasI
dt
．ThebasicLBPhistogramofori ginal
imageIismappedontoI
dt
bymeansofhis togramspecification，
andtheresultedimageisdenot edasI
LBP
．Thefusionimageof
I
dt
an dI
LBP
isI
pro
．Multi-scalespecialdeco mpositionisapplied
toI
pro
．Thecombinatio nofeachscalerotationinvariantLBP
histogramsisus edasfeatureforrecognition．Experimentsshow
thatt hemethodpresentedinthispaperhashigherrecognitionrateandefficiency．
1TheQuarterSampleShift（Q-s hift）
DT-CWT
Thediscretewavelettransform（DWT ）ismostcommonly
usedinitsmaximallydecimatedform ．Thisworkswellfor
compressionbutitsusesforother signalanalysisand
reconstructiontaskshavebeenha mperedbytwomain
disadvantages：lackofshiftinvari anceandpoordirectional
selectivity．InRefs．［6，7］，Kingsburyintroducedanewform
ofDWT，whichgenerat edcomplexcoefficientsbyusingadual-
treeofwavele tfilterstoobtaintheirrealandimaginaryparts．
Thi sintroduceslimitedredundancyandallowsthetransformt o
provideapproximateshiftinvarianceanddirection allyselective
filterswhilepreservingtheusualpro pertiesofperfect
reconstructionandcomputational efficiencywithgoodwell-
balancedfrequencyrespon ses．TheDT-CWThasreducedover
completenesscompare dwiththeshiftinvariantDWT（SIDWT），
anincreaseddi rectionalsensitivityovertheDWTthatisableto
dist inguishbetweenpositiveandnegativeorientationsgivin gsix
distinctsub-bandsateachlevel，theorientatio nsofwhichare
±15°，±45°，±75°．TheDT-CWTgivesperfe ct
reconstructionasthefiltersarechosenfromaperf ect
reconstructionbi-orthogonalset．TheQ- shiftDT-CWTisa
variantoftheearlierform，inordert ogivethedual-tree
improvedorthogonalityandsymme tryproperties．TheQ-shift
versionoftheDT-CWTissh owninFig．1，inwhichallthe
filtersbeyondlevel1are evenlength，buttheyarenolonger
strictlylinearpha se．Insteadtheyaredesignedtohaveagroup
Receivedd ate：2011-09-28
*Correspondenceshouldbeaddressed toCHENLei，E-mail：chenlei@suda．edu．cn

72 JournalofDonghuaUniversity（Eng．Ed．）Vol．29，No．1（ 2012）
delayofapproximately14sample（+q）．Therequi reddelay
differenceof12sample（2q）isthenachieved byusingthetime
reverseofthetreeafiltersintreebs othatthedelaybecomes
3q．Furthermore，thefilterco efficientsarenolonger
symmetric，anditisnowpossi bletodesigntheperfect-
reconstructionfiltersets tobeorthonormal，sothatthe
reconstructionfilters arejustthetimereverseoftheequivalent
analysisfi ltersinbothtrees．Henceallfiltersbeyondlevel1are
derivedfromthesameorthonormalprototypeset．
Fig ．2Q-shiftDT-CWTonafacialexpressionimage：（a）origina l
image；（b）realpart；（c）imaginarypart；（d）magnitu de
2
2．1
RotationInvariantLBPsandFacialExpressionFeatureExtraction
LBPs
TheLBPope ratorwasfirstintroducedbyOjalaetal．
［8］
andw asprovedapowerfulmeansoftexturedescription．The
operatorlabelsthepixelsofanimagebythresholdinga3×3
neighborhoodofeachpixelwiththecentervalueandconsideringtheresultsasabinarynumber（seeFig．3for an
illustration）．Bydefinition，LBPoperatordiscar dsthe
illuminationchanges，sinceitdependsonthegr ay-scale．This
makesitattractivesincedealingwith varyingillumination
whereinitisthemainconcernin ourresearch．
Fig．1TheQ-shiftDT-CWT，givingrealan dimaginarypartsof
complexcoefficientsfromtreeaa ndtreebrespectively
（q=14sampleperiod）
Notet hatfortheQ-shiftCWTeachcomplexwaveletbasis
isce nteredontheequivalentcomplexscalingfunctionbasis， andeachoftheseiscenteredbetweenapairofadjacentc omplex
basesfromtheprevious（finer）level．Inthisw ay，each
complexwaveletcoefficientatlevelkhastwo complexchildren
locatedsymmetricallyaboveitatle velk－1．Fortheoddeven
DT-CWT，suchsymmetriesdonot occur．
Insummary，Q-shiftDT-CWThasthefollowingproperties：approximateshiftinvariance，gooddirect ional
D）withgabor-likefilters（alsotrueselectivi tyin2-dimension（2-
forhigherdimensionality，m-D），perfectreconstructionusing
Nshortlinear-phasef ilters，limitedredundancy，efficientorder-
comput ation，improvedorthogonalityandsymmetryproperties． TheQ-shiftDT-CWTonafacialexpressionimageisshown in
Fig．2．
Fig．3ThebasicLBPoperator（LBP=1+8+3 2+128=169）
LBPisnotrotationinvariant，whichisund esirablein
certainapplications．Itispossibletode finerotationinvariant
versionsofLBP，andonesolut ionisillustratedinFig．4，where
LBPROTrepresentst hevalueofonerotationinvariantpattern．
Thebinary valuesofthethresholdedneighborhoodaremapped
int oan8-bitwordinclockwiseorcounter-clockwiseorder．An
arbitrarynumberofbinaryshiftsisthenmade，untilt heword
matchesoneofthe36differentpatternsof“0”a nd“1”an8-bit
wordcanformunderrotation．Theindexo fthematching
patternisusedasthefeaturevalue，des cribingtherotation
invariantLBPofthisparticular neighborhood
［9］
．
Fig．4Rotation-invarian tversionofLBP
Thederivedbinarynumberscodifyloca lprimitives
includingdifferenttypesofcurvededge s，spots，flatareas，etc．
（asshowninFig．5），soeachL BPcodecanberegardedasa
micro-pattern．Thelimitat ionofthebasicLBPoperatorisits
small3×3neighborh oodwhichcannotcapturedominant
featureswithlarge scalestructures．Hencetheoperatorwaslater
extend edtouseneighborhoodofdifferentsizes
［10］
．Us ing
circularneighborhoodsandbilinearlyinterpola tingthepixel
valuesallowanyradiusandnumberofpix elsinthe
neighborhood．Figure6illustratessomeexa mplesofthe
extendedLBPoperator，wherethenotation （P，R）denotesa
neighborhoodofPequallyspacedsampl ingpointsonacircleof
radiusofRthatformacircular lysymmetricneighborset．

JournalofDonghua University（Eng．Ed．）Vol．29，No．1（2012）
73
TheL BPoperatorLBP
（P，R）
produces2
P
differ entoutput
values，correspondingtothe2
P
di fferentbinarypatternsthatcan
beformedbythePpixe lsintheneighborset．Ithasbeenshown
thatcertainpa tternscontainmoreinformationthanothers．
Therefo re，itispossibletouseonlyasubsetofthe2
P
LBPs to
describethetextureofimages．Ojalaetal．
［8］
calledthese
fundamentalpatternsasuniformpat terns．AnLBPiscalled
uniformifitcontainsatmosttw obitwisetransitionsfrom0to1
orviceversawhentheb inarystringisconsideredcircular．For
example
，00000000，001110000，and11100001areuniform
patte rns．Itisobservedthatuniformpatternsaccountfornearl y
90%ofallpatternsinthe（8，1）neighborhoodandfora bout
70%inthe（16，2）neighborhoodintextureimages．
Accumulatingthepatternswhichhavemorethan2trans itions
2
intoasinglebinyieldsanLBPoperator，d enotedLBP
riu
（P，R）
，
withlessthan2
P
bins．Superscriptriu2standsforrotation
inv ariantuniformLBPandlabelingallremainingpatternswit ha
singlelabel
［11，12］
．Forexample，thenum beroflabelsfora
neighborhoodof8pixelsis256forth estandardLBPbut59for
riu2
LBP
16，2
．Af terlabelinganimagewiththeLBPoperator，a
histogra mofthelabeledimagef
l
（x，y）canbedefinedas
H
i
=
andhighfrequencycomponents（cDjs），w ithQ-shiftfilter．
（2）ZeroallthecoefficientsincA j，andpreservethe
coefficientincDjs．Employinvers eofDT-CWTbyusing
modifiedcAjtogetherwithcDjs，re ferredasI
dt
．
（3）DerivethebasicLBPhistog ramfromtheoriginal
imageI．
（4）MapthebasicLBP histogramofIontoI
dt
，bymeans
ofhistogram specification．Theresultedimageisdenotedas
I
LBP
．
（5）ConvertbothI
dt
andI
LBPintofrequencydomainbyQ-
shiftDT-CWTandthenfus eapproximationanddetails
coefficientsrespective ly．AnimageI
pro
isreconstructedbythe
fuse dcoefficients．
（6）ThepreprocessedimageI
pro isdividedintomulti-level
sub-regions．Therota tioninvariantLBPhistogramsarederived
fromtheses ub-regions，normalizeddependingontheregion
sizes ，andweightedaccordingtotheregionlocation．These
histogramsbuiltfromsub-blockscaneffectivelydescrib efacial
expressionmicro-patterns．Theyarecombine dandservedas
featurevectorsforrecognition．
F igure7showslowandhighfrequencydirectional
coeff icientsofafacialexpressionimage．
Fig．7
∑
I（f
（x，y）
x，y
l
=i），i=0，1，…，n－1，（
1 ）
wherenisthenumberofdifferentlabelsproducedbyt heLBP
operatorand
（2）
I（A）=
1，Aistrue，
0，Aisfalse．
Approximationanddetailscoeffici ents，withQ-shiftfilter：
（a）anoriginalimage；（b）a pproximationcomponent；
（c）sixdirectionaldetails components（magnitude）at
level2；（d）detailscompon entsatlevel3
{
ThisLBPhistogramcontainsinfor mationaboutthe
distributionofthelocalmicro-patt erns，suchasedges，spotsand
flatareas，overthewhol eimage，soitcanbeusedto
statisticallydescribeima gecharacteristics．
Aszeroingallthelowfrequencyc oefficientsandpreserving
thehighfrequencycoeffi cients，wecangetthereconstructed
imageFig．8（b）by employinginverseofDT-CWT．Figure
8（a）isthebasicL BPhistogramofFig．7（a），whichismapped
ontoFig．8（b ），andFig．8（c）istheresultedimagebymeans
ofhistog ramspecification．Figure8（d）isthefusionimageof
F ig．8（b）andFig．8（c）byapplyingDT-CWTfusionmethod．
2．2Facialexpressionfeatrueextraction
Automatic facialexpressionrecognitioninvolvestwovital
asp ects：facialrepresentationandclassifierdesign．Facia l
representationistoderiveasetoffeaturesfromori ginalface
imagestoeffectivelyrepresentfaces．The optimalfeatures
shouldminimizewithin-classvaria tionsofexpressionswhile
maximizebetweenclassvar iations．Ifinadequatefeaturesare
used，eventhebes tclassifiercouldfailtoachieveaccurate
recogniti on．Inthispaperafeatureextractionalgorithmbased
ontheDT-CWTandLBPhistogramsisproposed．Thewhole
processoffeatureextractionalgorithmiscarriedoutasf ollows．
（1）Multi-levelDT-CWTisemployedtoconvert thegray-
scalefacialimageIintotwocomponents，low frequency（cAj）

74
JournalofDonghuaUni versity（Eng．Ed．）Vol．29，No．1（2012）
recognitionpe rformanceandfeaturevectorlength．Thusface
images aretotallydividedinto32（1+2+4+9+16=32）
regions． TheLBPfeaturesextractedfromeachsub- regionare
normalizedaccordingtothesub-regionsiz esandthen
concatenatedintoasinglefeaturehistogr amwiththelengthof
416（32×26=416）．
3
Fig．8 Imagefusion：（a）thebasicLBPhistogramofFig．7（a）；
（b）theimagereconstructedbyhighfrequencycoefficient s；
（c）resultedimagebymeansofhistogramspecificat ion；
（d）fusedimageforfeatureextraction
Exper imentalResults
Thecharacteristicvectorsareextra ctedusingthemulti-level
LBPhistogramsoftheprepr ocessedimageFig．8（d）．AnLBP
histogramcomputedove rthewholefaceimageencodesonlythe
occurrencesoft hemicro-patternswithoutanyindicationabout
their locations．Toalsoconsidershapeinformationoffacialexpression，faceimagesareequallydividedintosmallr egions
R
0
，R
1
，…，R
m
toextr actLBPhistograms（asshowninFig．9
（a））．TheLBPfeat uresextractedfromeachsub-regionare
Forexperimen tsweusedimagesfromoneofthepopular
databasesforf acialexpressionrecognition，theJAFFEdatabase．
Th eJAFFEdatabasecontains213imagesofsevenfacial
ex pressions（6basicfacialexpressionsand1neutral）posed by
10Japanesefemalemodels．Foreachwoman，thereare 2-4
imagesofeveryfacialexpression．EachimageisaT IFFimage
withsize256×256and256graylevels．Allthe imagesare
takenagainstahomogeneousbackgroundwit hthesubjectsin
frontalposition．Someimagesof6bas icfacialexpressions
（anger，disgust，fear，joy，sad ness，andsurprise）areshownin
Fig．10andcorrespond ingimagesprocessedbyQ-shiftDT-CWT
inFig．11．
Fig．9ThenormalizedLBPhistograms：（a）sub-regionsatea chlevel，
riu2
（b）labeledimageofmicro-pattern s，（c）LBP
24，3
normalized
histogramsof9sub -regions，（d）concatenatedhistogramserved
asfeatu reforrecognition
concatenatedintoasingle，spatia llyenhancedfeature
histogram．Theextractedfeatur ehistogramrepresentsthelocal
textureandglobalsh apeoffaceimages．Someparameterscan
beoptimizedfo rbetterfeatureextraction．OneistheLBP
operator，a ndtheotheristhenumberofregionsdivided．We
riu2selecttheLBP
24，3
operator，bywhichwedefine 26rotation
invariantmicropatterns，anddividethef aceimagesinto1，2，
4，9，16regionsrespectively，giv ingagoodtrade-offbetween
Thecorrectlabelsofthet rainingsamplesarevery
importantforrecognition．A sseveralexpressionimagesare
markedwronginthedat abase，wecorrectthembeforeour
experiment．Wedodif ferentexperimentsusingdifferent
characteristics andtwomatchingmethodstoanalyzetheface
recogniti onperformances．ThecomposedLBPhistogramsof32
mul ti-scalesub-regionsareservedasfeaturesandthose
histogramsofsub-regionsinvolvingmouthoreyesaresetb igger
weight．Thiscanimprovetherecognitionaccura cyeffectively．
Weadoptedtemplatematchingtoclass ifyfacialexpressionsand
employedtwomethodsforsi milaritymeasure．
（1）Foreachexpressionofonesubje ct，wetestthreetimes
inturnandtaketheaveragereco gnitionrateasthefinalresult．
Wetakeoneofthefaci alexpressionimagesasatestsampleand
therestastra iningoneseverytime．Thereisnooverlapbetween
thet rainingandtestimages．WeemployEuropeandistance
m easurementforrecognition．

JournalofDongh uaUniversity（Eng．Ed．）Vol．29，No．1（2012）
75
（2 ）Intraining，thehistogramsofexpressionimagesina
givenclassareaveragedtogenerateatemplateforthiscla ss．A
nearest-neighborclassifierisusedasdissimil aritymetricfor
comparingatargetfacehistogramtot hemodelhistogram．
Resultsobtainedfromthediffere ntexperimentsare
presentedinTable1．Inthetablewe canseehowdifferent
featuresandsimilaritymeasure saffectrecognitionrate．
Table1Recognitionrateso btained
Recognitionrate
（6expressions）
65 %
84．5%
89%
100%
Features
matchingm ethod
LBPof16same-scalesub-regionsdistance
b etweentestingsampleandtrainingsamples
LBPof32mu lti-scalesub-regionsdistance
betweentestingsamp leandtrainingsamples
WeigtedLBPof32multi- scalesub-regions
distancebetweentestandtraining samples
WeigtedLBPof32multi-scalesub-regions
distancebetweentestsampleandclasscentre
weused atemplatematchingtoclassifyfacialexpressionsforits
simplicity．Comparedtherecognitionresultsobtain edwithour
facialfeaturestothoseobtainedwithPCAa ndLDAapproaches
（76．3%and69．5%forPCAandLDA，resp ectively），the
methodproposedinthispaperclearlys howedthebetter
performanceinrecognitionefficien cyandaccuracy．
References
［1］MehrabianA．Sile ntMessages［M］．WadsworthPublishing
1971．Company， Inc．，Belmont，CA，
［2］EkmanP，FriensenW．FacialActi onCodingSystem（FACS）：a
TechniquefortheMeasureme ntofFacialMovement［M］．Palo
Alto：ConsultingPsych ologistsPress，1978．
［3］EkmanP，FriensenW．Unmaski ngtheFace［M］．PaloAlto：
ConsultingPsychologistsP ress，1984．
［4］PanticM，RothkrantzLJM．FacialActio nRecognitionforFacial
．IEEEExpressionAnalysisfr omStaticFaceImages［J］
TransactionsonSystems，Man ，andCybernetics，2004，34（3）：
1449-1461．
［5］Pa nticM，RothkrantzL．AutomaticAnalysisofFacial
Exp ressions：theStateofArt［J］．IEEETransactionsonPatter n
2000，22（12）：1424-1445．AnalysisandMachineIntel ligence，
［6］KingsburyNG．TheDual-TreeComplexWave letTransform：a
NewEfficientToolforImageRestorat ionandEnhancement［C］．
ProceedingsinEUSIPCO98，Rh odes，Greece，1998：319-322．
［7］KingsburyNG．Comple xWaveletsforShiftInvariantAnalysis
．JournalofAp pliedandandFilteringofSignals［J］
ComputationalH armonicAnalysis，2001，10（3）：234-253．
［8］OjalaT，P ietikinenM，HarwoodD．AComparativeStudyof
Textur eMeasureswithClassificationBasedonFeature
Distr ibutions［J］．PatternRecognition，1996，29（1）：51-59．［9］PietikinenM，OjalaT，XuZ．Rotation-InvariantTex ture
ClassificationUsingFeatureDistributions［J］．Pattern
Recognition，2000，33（1）：43-52．
［10］O jalaT，PietikinenM，MenpT．MultiresolutionGray-Sc ale
andRotationInvariantTextureClassificationwi thLocalBinary
Patterns［J］．IEEETransactionsonPat ternAnalysisandMachine
2002，24（7）：971-987．Intel ligence，
［11］AhonenT，HadidA，PietikinenM．FaceRe cognitionwithLocal
BinaryPatterns［C］．Proceeding softhe8thEuropeanConference
onComputerVision，Pr ague，TheCzechRepublic，2004：469-
481．
［12］Moo reS，BowdenR．LocalBinaryPatternsforMulti-viewFacial
ExpressionRecognition［J］．ComputerVisionandImag e
Understanding，2011，115（4）：541-558．
4Conclu sions
InthispaperweanalyzedthebasicprincipleofD T-CWT
andLBPs．AnovelmethodbasedonQ-shiftDT-CWTa nd
rotationinvariantLBPwasproposedwhichwaseffic ientfor
recognition．Q-shiftDT-CWTwasusedtoresol veillumination
variationinexpressionverificatio n．RotationinvariantLBPs
werecapableofdescribing textureandshapeinformation．To
enhancethefacialr epresentation，wedividedthepreprocessed
facialim agesintoseveralsub-regionsofdifferentscales．The
LBPhistogramsofsub-regionswerenormalizeddepending on
theregionsizesandgaveadifferentweightdependi ngonthe
roleofthegivenregionsinrecognition．Fori nstance，sincethe
mouthregionswereimportantforre cognition，ahighweight
couldbeattributedtothecor respondingLBPhistograms．The
combinedhistogramst hatcouldeffectivelydescribefacial
expressionmic ro-patternswereservedasfeaturesfor
recognition． Ourmaingoalinthispaperwastoshowthehigh
discrimi nativepoweroftheproposedfacialfeatures．Therefore，