ADAPTIVEIMAGESEGMENTATIONBASEDONCOLORANDTEXTURE
JunqingChena0,ThrasyvoulosN.Pappasa1
ElectricalandComputerEngineeringDept.
NorthwesternUniversity,Evanston,IL60208
AleksandraMojsilovic,BerniceRogowitz
IBMT.J.WatsonResearchCenter
Hawthorn,NY10532
ABSTRACT
Weproposeanimagesegmentationalgorithmthatisbasedonspa-
tiallyadaptivecolorandtexturefeatures.Thefeaturesarefirst
developedindependently,andthencombinedtoobtainanoverall
segmentation.Texturefeatureestimationrequiresafiniteneigh-
borhoodwhichlimitsthespatialresolutionoftexturesegmenta-
tion,whilecolorsegmentationprovidesaccurateandpreciseedge
localization.Wecombineapreviouslyproposedadaptivecluster-
ingalgorithmforcolorsegmentationwithasimplebuteffective
texturesegmentationapproachtoobtainanoverallimagesegmen-
tation.Ourfocusisinthedomainofphotographicimageswith
anessentiallyunlimitedrangeoftopics.Theimagesareassumed
tobeofrelativelylowresolutionandmaybedegradedorcom-
pressed.
1.INTRODUCTION
ThefieldofContent-BasedImageRetrieval(CBIR)hasmadesig-
nificantadvancesduringthepastdecade[1,2].Anumberofprac-
ticalsystemshavebeenproposed,andtheMPEG-7standardspec-
ifiesdescriptorsforvisualcontent[3].Oneofthemostchalleng-
ingproblemsisimagesegmentation.Whilesignificantprogress
hasbeenmadeintexturesegmentation(e.g.,[4–7])andcolorseg-
mentation(e.g.,[8–10])separately,thecombinedtextureandcolor
segmentationproblemisconsiderablymorechallenging[11–13].
Inthispaper,weproposeanimagesegmentationalgorithm
thatisbasedonspatiallyadaptivecolorandtexturefeatures.These
featuresarebasedonperceptualmodelsandprinciplesaboutthe
processingoftextureandcolorinformation.Thecolorandtexture
featuresarefirstdevelopedindependently,andthencombinedto
obtainanoverallsegmentation.Oneofthemostsignificantdiffer-
encesbetweenthecolorandtexturefeaturesisthattexturerequires
afiniteneighborhoodtobedefined.Thislimitstheresolutionof
texturesegmentation.Colorsegmentation,ontheotherhand,can
provideveryaccurateandpreciseedgelocalization.Toaccom-
plishthisgoal,ourcolorsegmentationisbasedontheadaptive
clusteringalgorithm[8].Fortextureanalysis,weuseanestimate
oftheenergyofthecoefficientsofawaveletdecomposition[14].
Thekeytotheproposedapproachistheuseofthemedianoperator
forestimatingthetextureenergyinawindowaroundeachpixel.
Theadvantageofthemedianisthatitrespondstotexturewithin
uniformregionsandsuppressestexturesassociatedwithtransi-
tionsbetweenregions.Finally,inordertocombinethetexture
andcolorinformation,westartwiththetexturesegmentationand
usethecolorsegmentationtorefineit.
Ourfocusisinthedomainofphotographicimages,butwith
anessentiallyunlimitedrangeoftopics(people,nature,buildings,
textures,objects,indoorscenes,etc.).Theimagesareassumedto
a2SummerinternatIBM,2001
a3ConsultantatIBM,September2001
beofrelativelylowresolution(e.g.,a4a6a5a7a5a9a8a10a4a11a5a7a5)andoccasionally
degradedorcompressed.Anaddedadvantageisthatimageaccess
andprocessingtimewillbesignificantlyreduced.
Theimagesegmentationresultscanbeusedtoderiveregion-
widecolorandtexturefeatures,whichinturn,togetherwiththe
segmentlocation,boundaryshape,andregionsize,canbeusedto
extractsemanticinformation.Akeytothesuccessoftheproposed
approachistherecognitionofthefactthatitisnotnecessaryto
obtainacompleteunderstandingofagivenimage:Inmanycases,
itisenoughtoidentifyafewfeatures(colorcomposition,presence
ofkeysegmentssuchas“sky,”“mountains,”people,”etc.)tobe
abletoclassifyitinagivencategory[15].Thus,someofthe
regionswillbeclassifiedas“complex”or“noneoftheabove,”
andassuchwillstillplayasignificantroleinsceneanalysis.
InSection2,wediscusscolorfeatureextraction.Ourap-
proachfortexturefeatureextractionispresentedinSection3.Sec-
tion4discussesthecombinationoftextureandcolorfeaturesto
obtainanoverallsegmentation.
2.COLORFEATUREEXTRACTION
Colorhasbeenusedextensivelyasalow-levelfeatureforimage
retrieval[1,3].Manyoftheexistingtechniquesarebasedonthe
colorimagehistogram.Eventhoughsuchtechniqueshavebeen
quitesuccessfulingivensettings,theyhavesomenotableshort-
comings.First,thehistogramdoesnotincorporateanyspatialin-
formation.Second,thecolorhistogramistoofinelyquantizedin
colorspace,andhence,doesnottakeintoconsiderationthefact
thatthehumanvisualsystemcanonlyperceiveafewcolorsata
time.Theproposedapproachattemptstoobtainlow-levelcolor
featuresthatincorporateknowledgeofhumanperception.
Oneofthemostimportantcharacteristicsofhumancolorper-
ceptionisthatthehumaneyecannotsimultaneouslyperceivea
largenumberofcolors[16].Moreover,thenumberofcolorsthat
canbeinternallyrepresentedandidentifiedincognitivespaceis
about30[17].Inadditiontoprovidingaveryefficientrepresen-
tation,thecompactsetofcolorcategoriesalsomakesiteasierto
captureinvariantpropertiesinobjectappearance[18].
Maetal.[11]werethefirsttoproposeacompactcolorrepre-
sentationintermsofdominantcolorsforimagesegmentationand
retrieval.Therepresentationtheyproposedconsistsofthedom-
inantcolorsandthecorrespondingpercentageofoccurrenceof
eachcolor:a12
a13a15a14a17a16a19a18a21a20a23a22a25a24a21a26a27a22a29a28a30a24a25a31a32a14a34a33a7a24a36a35a23a35a37a35a38a24a40a39a41a24a42a26a27a22a44a43a46a45
a5
a24a36a33a37a47a42a48(1)
whereeachofthedominantcolors,a20a37a22,isathreedimensionalvec-
torinRGBspace,anda26a49a22arethecorrespondingpercentages.
Thereareanumberofapproachesforextractingthedominant
colors.Oneapproachistousevectorquantization(VQ)toobtain
asetofcolorswhichminimizethemean-squarequantizationer-
rorforallthepixelsoftheimagesinagivendatabase[11,16].
Thisrequiresalargetrainingset,whichdramaticallyincreases
thecomputationalburdenforthecodebookdesign.Analternative
techniquethatavoidsthisproblemwasproposedbyMojsilovicet
al.[19].Ratherthantryingtofindasetofcolorsthatisrepresen-
tativeofallimagesoraparticularimagedatabase,onecanobtain
thedominantcolorsofagivenimage.Forexample,onecanuse
VQonasingleimage.Whiletheresultingcolorsmaybeuseful
incharacterizingtheimageasawhole,theresultingsegmentation
couldbequiteinadequateduetolackofspatialconstraintsand
spatialadaptation[8].
Theabovedominantcolorextractiontechniquesdependonthe
assumptionthatthecharacteristiccolorsofanimagearerelatively
constant,i.e.,theydonotchangeduetovariationsinillumination,
perspective,etc.Thisistrueforimagesoffabrics,carpets,interior
designpatterns,andotherpuretexturesthatweretheprimaryfocus
oftheworkpresentedin[16].Ourfocusofattention,however,is
onamoregeneralclassofimagesthatincludesoutdoorandindoor
scenes,includinglandscapes,cityscapes,plants,animals,people,
andman-madeobjects.Tohandlesuchimages,onehastoaccount
forcolorandlightingvariationsinthescene.
Arelativelysimpleandquiteeffectivealgorithmthatonecan
useforobtainingthedominantcolorsofanimageinthiswider
classofimagesisthecolorsegmentationalgorithmproposedby
ComaniciuandMeer[10].Itisbasedonthe“meanshift”algo-
rithmforestimatingdensitygradients,andessentiallyworkswith
theimagehistogram,eventhoughitalsoattemptstoincorporate
spatialconstraintsbyimposingconstraintsontheconnectivityof
thedetectedregions.However,itdoesnottakeintoconsideration
thatthecolorsinanimage(andhencethedominantcolors)may
beslowlyvaryingacrosstheimage.
Instead,weproposetousespatiallyadaptivedominantcol-
ors.Thisisnecessitatedbythespatiallyvaryingimagecharac-
teristicsandtheadaptivenatureofthehumanvisualsystem.For
example,anobserver’snotionofablueorbrownorgreencolor
ishighlydependentonthesurroundingcolors;moreover,itvaries
withthelightingconditionsandthecolorsofthedisplaydevice.
Thespatiallyadaptivedominantcolorscanbeobtainedbythe
adaptiveclusteringalgorithm(ACA)proposedin[8]andextended
tocolorin[9].TheACAisaniterativealgorithmthatusesspa-
tialconstraintsintheformofMarkovrandomfields(MRF).The
initialestimateisobtainedbythea50-meansalgorithm,whiches-
timatestheclustercenters(i.e.,thedominantcolors)byaveraging
thecolorsofthepixelsineachclassoverthewholeimage.As
thealgorithmprogresses,thedominantcolorsareupdatedbyav-
eragingoveraslidingwindowwhosesizeprogressivelydecreases.
Thus,thealgorithmstartswithglobalestimatesandslowlyadapts
tothelocalcharacteristicsofeachregion.
Whilethelocalcoloradaptationisnotsuitedforcomparing
wholeimages,weshowthatitsignificantlyimprovesimageseg-
mentation,asitvirtuallyeliminatesfalsecontoursandprovides
preciseboundarylocation.Moreover,incontrasttotheotherap-
proaches,theACAisquiterobusttothenumberofclasses.This
isbecausethecharacteristiclevelsofeachclassadapttothelocal
characteristicsoftheimage,andthus,regionsofentirelydifferent
intensitiescanbelongtothesameclass,aslongastheyaresepa-
ratedinspace.Wefoundthatthebestchoiceisfourclasses[8].
Fig.1comparestheComaniciu-Meeralgorithm[10](using
the“oversegmentation”settingfordeterminingthenumberofdom-
inantcolors)totheACA.Theimageresolutionisa51a7a52a11a53a9a54a55a51a19a56a37a57pix-
els.NotethefalsecontoursintheComaniciu-Meeralgorithmin
thewaterandthesky.Also,whiletherearecolorvariationsinthe
forestregion,thesegmentboundariesdonotappeartocorrespond
toanytruecolorboundaries.TheACAontheotherhand,smooths
overthewater,sky,andforestregions,whilecapturingthedom-
inantedgesofthescene.NotethattheACAwasdevelopedfor
imagesofobjectswithsmoothsurfacesandnotexture.Forsome
texturedregions,likethemountainarea,theACAoversegments
theimage,butthesegmentsdocorrespondtoactualtexturede-
tails.Thus,weneedsomeothermechanismtoconsolidatesuch
smallsegmentsintoregions.
ColorFeatures:TheACAprovidesasegmentationoftheimage
intoclasses.IntheexampleofFig.1(c),eachpixelwaspainted
withtheaveragecolorofthepixelsinitsneighborhoodthatbe-
longtothesameclass[8].Assumingthatthedominantcolorsare
slowlyvarying,wecanassumethattheyareapproximatelycon-
stantintheimmediatevicinityofapixel.Thus,wecancountthe
numberofpixelsineachclasswithinagivenwindow,andaverage
theircolorvaluestoobtainafeaturevectorthatconsistsofafew
(uptofour)dominantcolorsandtheassociatedpercentages.
ColorMetric:Oncewehavea(spatiallyvarying)representation
oftheform(1),weneedametricthatmeasurestheperceptualsim-
ilaritybetweenthefeaturevectors.Basedonhumanperception,
thecolorcompositionoftwoimagesorimagesegmentsissimi-
lar,iftwoconditionsaresatisfied[16,19]:Thecolorsaresimilar,
andthecorrespondingareapercentagesaresimilar.Thedefinition
ofametricthattakesintoaccountboththecolorandareadiffer-
ences,dependsonthemappingbetweenthedominantcolorsof
thetwoimages[19].Varioussuboptimalsolutionshavebeenpro-
posed[11,16].Mojsilovicetal.[19]definedametricthatfindsthe
optimalmappingbetweenthedominantcolorsoftheimagesand
callitthe“optimalcolorcompositiondistance(OCCD).”Ingen-
eral,thismetricrequiresmorecomputationthansimplermetrics.
However,sinceweareprimarilyinterestedincomparingimage
segmentsthatcontainonlyafewcolors(atmostfour),theaddi-
tionaloverheadfortheOCCDisreasonable.Moreover,weusea
moreefficientimplementation.
Theproposedcolorfeaturevectorandassociatedmetricincor-
porateknowledgeofthehumanvisualcharacteristics.Themost
importantnewelementoftheproposedapproachisthefactthat
thefeaturevectorconsistsofasmallnumberofspatiallyvarying
dominantcolors.Thisvariationreflectsthespatiallyvaryingimage
characteristicsandtheadaptivenatureofthehumanvisualsystem.
3.TEXTUREFEATUREEXTRACTION
Inthissection,wetrytoisolatethetexturalfeatureextractionfrom
thatofcolor.Wethusconsideronlythegrayscalecomponentof
theimagetoobtainthetexturefeaturesthatcanbeusedtoobtain
anintermediatesegmentation,whichcanthenbecombinedwith
thecolorfeaturestoproducethefinalsegmentation.Thisisin
contrasttosomeotherapproaches[13,16]wherethecolorquan-
tization/segmentationisusedtoobtainanachromaticpatternmap
whichisthebasisfortexturefeatureextraction.
Likemanyoftheexistingalgorithmsfortextureanalysisand
synthesis(e.g.,[6,20],ourapproachisbasedonamultiscalefre-
quencydecomposition.Suchdecompositionshavebeenwidely
usedasdescriptionsofearlyvisualprocessinginmammalsand
havealsobeenusedasthebasisfortextureclassificationandseg-
mentation(e.g.,[21–23]).
Whilethetexturesynthesisproblemrequiresaprecisemodel
inordertoaccuratelysynthesizeawiderangeoftextures,the
(a)OriginalColorImage(b)Comaniciu-MeerAlgorithm(c)AdaptiveClusteringAlgorithm(ACA)
Fig.1.ColorImageSegmentation(allimagesshownincolor)
modelforthesegmentationproblemcanbequitecrude.More-
over,wewanttocapturesuchinformationfromrelativelysmall
images,e.g.,a58a6a59a7a59a61a60a62a58a6a59a7a59pixels,whichmayhavebeencompressed
ordegraded.Thus,weexpectourtexturemodelstobealotsimpler
thanthoserequiredfortexturesynthesis.
Oneofthekeyideasintheproposedapproachisthatwedo
notneedtorecognizeeveryregioninanimage.Largeareasofan
imagemaybeclassifiedas“toocomplex,”or“nonoftheabove,”
andassuchplayaroleinoverallimageclassification.Inthispaper,
weusesimpletexturefeaturestoobtainasegmentationinalimited
numberoftextureclasses(“smooth,”“horizontal,”“vertical,”and
“complex”).Inthenextsection,weshowthatthissimpletexture
classificationcanleadtoimpressiveresults.
Forourtextureanalysis,weusethe9/7biorthogonalwavelet
decomposition[14]whichisseparableandcomputationallyeffi-
cient.Sincetheimagesarefairlysmall(approx.a58a6a59a7a59a63a60a64a58a11a59a7a59pix-
els),weobtainaone-levelwaveletdecomposition,andusetheHL
andLHbands,whereHandLstandforthehigh-passandlow-
passbandineachofthehorizontalandverticalorientations.We
foundthatdiscardingtheHHbanddoesnotresultinanysignifi-
cantlossofvisualquality,andhenceitshouldnotbecriticalfor
textureanalysis.Wealsotriedthesteerablepyramid[24]buthave
notfoundanysignificantperformancedifferences.
Themostcommonlyusedfeaturefortextureanalysisinthe
waveletdomainistheenergyofthesubbandcoefficients[4,5].
Sincethecoefficientsarequitesparse,itisnecessarytoperform
sometypeofwindowoperationtoobtainamoreuniformchar-
acterizationoftexture.In[4,5],theaverageoftheenergyofthe
coefficientsinasmallwindowwasused.In[11,20]boththemean
andstandarddeviationofthemagnitudeoftheGabortransform
coefficientswereusedastexturefeatures.Inthispaper,weuse
themedianoftheenergyinawindow.Theadvantageoftheme-
dianisthatittendstofilterouttexturesassociatedwithtransitions
betweenregions.Insuchcases,theincreaseinwaveletcoeffi-
cientsduetotheregionboundaryisconcentratedalongtheedge
andisnotpickedupbythemedianoperator.Extensiveexperimen-
tationwiththeaverage,median,andmaximumoperatorsindicates
thattheaverageresultsinsignificantsmoothingandfalsetextures
alongtheregionboundaries,whilethemaximumprovidesunre-
liableresults.Themedianleadstosignificantlysuperiorresults.
Thesizeofthewindowmustbelargeenoughtocapturethelo-
caltexturecharacteristics,butnottoolargetoavoidbordereffects.
Wefoundthatfortheimageresolutionandviewingdistanceunder
consideration,aa65a66a60a62a65windowgivesthebestresults.
Finally,wetrieddifferentclusteringapproachestoobtainthe
(intermediate)texturesegmentation.Thesimplestandmostef-
fectivewastoapplytwo-levelK-meanstoeachofthehorizontal
andverticalcomponentsseparately.Oneoftheclustercenterswas
alwaysfixedat0(smoothtexture)andtheotherwasdetermined
bytheK-meansalgorithm.Theaddedadvantageofthisapproach
isthatweobtainfourtextureclasseswithobviousinterpretations:
smoothtexture(0,0),verticaltexture(1,0),horizontaltexture(0,1),
andcomplextexture(1,1).Fig.2(c)showstheresults,withsmooth
texturerepresentedbyblack,verticalbylightgray,horizontalby
darkgray,andcomplexbywhite.WealsoexperimentedwithVQ
(K-meansappliedtothevectorofthetwocoefficients),butfound
verylittledifferenceinperformanceformostimages.Thefinal
paperwillcontainamoreextensivecomparisonoffrequencyde-
compositionsandclusteringalgorithms.
4.COMBINATIONOFTEXTUREANDCOLOR
Wenowdiscussthecombinationofthetextureandcolorsegmen-
tationstoobtainthefinalimagesegmentation.Westartwiththe
texturesegmentationandusethecolorsegmentationtorefineit.
First,weconsiderthe“smooth”textureregions.Werelyonthe
colorsegmentationtodetermineiftheyshouldbefurthersubdi-
videdintosegmentsofdifferentcolor.Foreachconnectedsmooth
region,wefindalltheconnectedsegmentsthatbelongtodifferent
colorclassesandcomputetheaveragecolorofeachsegment.Re-
callthattheACAprovidesslowlyvaryingcolor.Wethenmerge
thesegmentswhosecolorsaresimilar.Thecomparisonisper-
formedinLabcolorspace;thethresholdissetat10%ofthedis-
tancerange.Fortheremainingsegments,wecomputetheaverage
colordifferenceacrossthecommonborder,andmergethemifitis
belowagiventhreshold.Finally,smallcolorregionsneighboring
non-smoothtextureregionsareconsideredinthenextstep.
Alltheothertexturedregions(“horizontal,”“vertical,”and
“complex”regions,aswellsmallborderregionsfromtheprevi-
ousstep)arethenconsideredtogether.Weapplyaregionmerging
algorithmbasedontwotexturefeatures:thegrayscaletextureclas-
sificationobtainedintheprevioussectionandthecolorfeatures
presentedinSection2.Thelatterconsistofthedominantcolors
andtheirfrequenciescomputedinaa58a68a67a9a60a69a58a68a67window.Foreach
pairofpixels,wecomputethecolorfeaturedifferenceaccording
totheOCCDcriteriondescribedinSection2,andmergethemif
itisbelowagiventhreshold.Thethresholdishigherifthepixels
belongtothesametextureclass,andloweriftheybelongtodif-
ferenttextureclasses.ResultsareshowninFig.2(d).Eachregion
(a)OriginalColorImage(b)ColorSegmentation(ACA)(c)TextureSegmentation
(d)CrudeSegmentation(e)FinalSegmentation(f)FinalSegmentation(onoriginalimage)
Fig.2.ColorandTextureImageSegmentation(a,b,d,e,fshownincolor)
inthefigureispaintedwithitsaveragecolor.
Finally,weadjustthesegmentbordersusingcolorinforma-
tion.Thatis,foreachpixelalongtheborderofthecrudeseg-
mentation(Fig.2(d),wecomparethecolorprovidedbytheACA
algorithmwiththeaveragecolorofthesegmentineithersideof
theborderandmoveittothesidewhoseaveragecoloritisclosest
to.Thissimpleapproachworksverywell,providedtheaverage
colorsofthetwoadjacentregionsarequitedifferent.Iftheaver-
agecolorsofthetworegionsareveryclose,thentheadjustmentis
notperformed.ThefinalresultisshowninFig.2(e)and(f).
5.REFERENCES
[1]Y.Rui,T.S.Huang,S.-F.Chang,“Imageretrieval:Cur-
renttechniques,promisingdirectionsandopenissues,”
J.Vis.Comm.Im.Repr.,v.10,p.39–62,Mar.1999.
[2]W.M.Smeulders,etal.,“Content-basedimageretrievalat
theendoftheearlyyears,”IEEETr.PAMI,v.22,p.1349–
1379,Dec.2000.
[3]B.S.Manjunath,etal.,“Colorandtexturedescriptors,”IEEE
Tr.CSVT,p.703–715,June2001.
[4]A.Kundu,J.-L.Chen,“TextureclassificationusingQMF
bank-basedsubbanddecomposition,”CVGIP,GMIP,v.54,
p.369–384,Sept.1992.
[5]T.Chang,C.-C.J.Kuo,“Textureanalysisandclassification
withtree-structuredwavelettransform,”IEEETr.IP,v.2,
p.429–441,Oct.1993.
[6]M.Unser,“Textureclassificationandsegmentationusing
waveletframes,”IEEETr.IP,v.4,p.1549–1560,Nov.1995.
[7]T.Randen,J.H.Husoy,“Texturesegmentationusingfilters
withoptimizedenergyseparation,”IEEETr.IP,v.8,p.571–
582,Apr.1999.
[8]T.N.Pappas,“Anadaptiveclusteringalgorithmforimage
segmentation,”IEEETr.SP,v.40,p.901–914,Apr.1992.
[9]M.M.Chang,M.I.Sezan,A.M.Tekalp,“AdaptiveBayesian
segmentationofcolorimages,”JEI,p.404–414,Oct.1994.
[10]D.Comaniciu,P.Meer,“Robustanalysisoffeaturespaces:
Colorimagesegmentation,”CVPR,June1997,p.750–755.
[11]W.Y.Ma,Y.Deng,B.S.Manjunath,“Toolsfortexture/color
basedsearchofimages,”HumanVisionandElectronicImag-
ingII,Feb.1997,Proc.SPIE,Vol.3016,p.496–507.
[12]S.Belongie,etal.,“Color-andtexture-basedimagesegmen-
tationusingEManditsapplicationtocontentbasedimage
retrieval,”ICCV,1998,p.675–682.
[13]Y.Deng,B.S.Manjunath,“Unsupervisedsegmentationof
color-textureregionsinimagesandvideo,”IEEETr.PAMI,
v.23,p.800–810,Aug.2001.
[14]A.Cohen,I.Daubechies,J.C.Feauveau,“Biorthogonal
basesofcompactlysupportedwavelets,”Commun.Pure
Appl.Math.,v.45,p.485–560,1992.
[15]A.Mojsilovic,B.Rogowitz,“Capturingimagesemantics
withlow-leveldescriptors,”ICIP,Oct.2001,p.18–21.
[16]A.Mojsilovi′c,etal.,“Matchingandretrievalbasedonthe
vocabularyandgrammarofcolorpatterns,”IEEETr.IP,v.1,
p.38–54,Jan.2000.
[17]G.Derefeldt,T.Swartling,“Colorconceptretrievalbyfree
colornaming,”Displays,v.16,p.69–77,1995.
[18]S.N.Yendrikhovskij,“Computingcolorcategories,”Hu-
manVisionandElectronicImagingV,Jan.2000,Proc.SPIE
Vol.3959.
[19]A.Mojsilovi′c,J.Hu,E.Soljanin,“Extractionofperceptu-
allyimportantcolorsandsimilaritymeasurementforimage
matching,retrieval,andanalysis,”IEEETr.IP,Toappear.
[20]B.S.Manjunath,W.Y.Ma,“Texturefeaturesforbrowsing
andretrievalofimagedata,”IEEETr.PAMI,v.18,p.837–
842,Aug.1996.
[21]J.R.Bergen,E.H.Adelson,“Visualtexturesegmentation
basedonenergymeasures,”JOSAA,v.3,p.99–,1986.
[22]M.R.Turner,“TexturediscriminationbyGaborfunctions,”
Biol.Cybern.,v.55,p.71–82,1986.
[23]J.Malik,P.Perona,“Preattentivetexturediscriminationwith
earlyvisionmechanisms,”JOSAA,v.7,p.923–932,1990.
[24]E.P.Simoncelli,W.T.Freeman,“Thesteerablepyramid:A
flexiblearchitectureformulti-scalederivativecomputation,”
ICIP,vol.III,Oct.1995,p.444–447.
|
|