La lecture à portée de main
Découvre YouScribe en t'inscrivant gratuitement
Je m'inscrisDécouvre YouScribe en t'inscrivant gratuitement
Je m'inscrisDescription
Sujets
Informations
Publié par | profil-zyan-2012 |
Nombre de lectures | 23 |
Langue | English |
Poids de l'ouvrage | 1 Mo |
Extrait
13335373931434547494
1.Introduction
accuratecandidacy,assegmentationshouldcomeupwith
partition(s)fromacandidatesegmentationset
[3]
.
GroupingisthediscoveryofintrinsicclustersindataWiththeadventofmediamakingiteasierandcheaperto
[1]
.Imagesegmentationisaparticularkindofgroupingin
collectandstoredigitalimages,unconstraineddigitalpho-
whichdataconsistsofanimage,andthetaskistoextracttographicimageshaveraisedthischallengeevenfurther
asregionstheobjectsausermayfindconceptuallydistincttowardsbothcomputationalefficiencyandrobustprocess-
fromeachother.Theautomationandoptimizationofthising.Considerforexamplethewell-knownbenchmarkimage
taskfacecomputationalissues
[2]
andanimportantcon-
lena
in
Fig.1
.Userswouldcertainlyconsiderthehatof
ceptualissue:basically,segmentationhasaccessonlytothethegirlasanobjectdifferentfromtheblurredbackground,
descriptionsofpixels(e.g.colorlevels)andtheirspatialre-andmostwouldconsiderhershoulderasdifferentfromher
lationships,whileauseralwaysuseshigherlevelofknowl-face.Nevertheless,duetothedistributionofcolors,itisvir-
edgetoclustertheimageobjects.Withoutsuchasignifi-tuallyimpossibleforsegmentationtechniquesbasedsolely
cantpriorworldknowledge,theaccuracyofgroupingisnotonlow-levelcues,suchasthecolors,tomakeacleansepa-
meanttobeoptimalityorevennear-optimality,butratherrationoftheseregions.Therightimagedisplaystheresult
ofouralgorithm.Theregionsfoundhavewhiteborders.
∗
Thisresultispresentedmoreindepthintheexperimental
Correspondingauthor.Tel.:+596727424;fax:+596727362.
section(Fig.
2
).Noticefromtheresultthesegmentationof
E-mailaddresses:
richard.nock@martinique.univ-ag.fr
(R.Nock),
frank.nielsen@acm.org
(F.Nielsen)
thehat,cleanlyseparatedfromthebackground,andalsothe
URLs:
http://www.univ-ag.fr/
∼
rnock
,
segmentationofthegirl’schin,whichisseparatedfromher
http://www.csl.sony.co.jp/person/nielsen/
.
shoulder.
0031-3203/$30.00
2004PatternRecognitionSociety.PublishedbyElsevierLtd.Allrightsreserved.
doi:10.1016/j.patcog.2004.11.009
911232527292
Abstract
Someauthorshaverecentlydevisedadaptationsofspectralgroupingalgorithmstointegratepriorknowledge,asconstrained
eigenvaluesproblems.Inthispaper,weimproveandadaptarecentstatisticalregionmergingapproachtothistask,asanon-
parametricmixturemodelestimationproblem.Theapproachappearstobeattractivebothforitstheoreticalbenefitsandits
experimentalresults,asslightbiasbringsdramaticimprovementsoverunbiasedapproachesonchallengingdigitalpictures.
2004PatternRecognitionSociety.PublishedbyElsevierLtd.Allrightsreserved.
Keywords:
Imagesegmentation;Semi-supervisedgrouping
9113151
www.elsevier.com/locate/patcog
PatternRecognition()–
Semi-supervisedstatisticalregionrefinementforcolorimage
segmentation
ba∗,RichardNock,FrankNielsen
a
GRIMAAG
-DépartementScientifiqueInterfacultaire,UniversitédesAntilles-Guyane,CampusdeSchoelcher,BP7209,97275Schoelcher,
Martinique,France
b
SonyComputerScienceLaboratories,Inc.,3-14-13HigashiGotanda,Shinagawa-Ku,Tokyo141-0022,Japan
Received9August2004
1357
SERPNIELCITRAlabolg:NACS--aydiV:NGAPabihtarP:DE1.0.5:REVDTD5912RP)lin:.gif.loc(21-1:PPMOC:EPYT.DORP
13579113115
2
RAITPCRL2E1I9N5RPESR.Nock,F.Nielsen/PatternRecognition()–
Fig.1.Image
lena
(left),andoursegmentation(right).Inthesegmentation’sresult,regionsfoundarewhitebordered(seetextfordetails).
Fig.2.Image
lena
(upperleft),anditssegmentationby
SRRB
,withoutbias(w/o),andwithbias(w/,see
Fig.1
).Inthesegmentations’
results,regionsfoundaredelimitedwithwhiteborders.Wehave
m
=
4
,
|
V
1
|=
2
,
|
V
2
|=
2
,
|
V
3
|=
2
,
|
V
4
|=
2.Theupperrighttabledisplays
someofthelargestregionsextractedfromthesegmentation(Reg.#X).Thebottomtableshowshow
lena
’sfaceisbuiltfromthetwo
modelswhosepixels,denotedbyredtrianglesontheupperrightimage,havebeenpointedbytheuser(eachmodel’scolorisrandom);the
numberbelow
×
2000is
SRRB
’siterationnumber.
CommongroupingalgorithmsforimagesegmentationOurapproachtosegmentation,whichgivestheresultsof
useaweightedneighborhoodgraphtoformulatethespa-
Fig.1
,isbasedonasegmentationframeworkpreviously
tialrelationshipsamongpixels
[1–6]
andthenformulatethe
studiedbyYuandShi
[1,7]
:groupingwithbias.Itispar-
segmentationasagraphpartitioningproblem.Anessentialticularlyusefulfordomainsinwhichtheusermayinteract
differencebetweenthesealgorithmsisthelocalityofthewiththesegmentation,byinputtingconstraintstobiasits
groupingprocess.ShiandMalik
[3]
andYuandShi
[1,7]
result:sensormodelsinMRF
[8]
,Human–computerinter-
solveitfromaglobalstandpoint,whereasFelzenszwalbandaction,spatialattentionandothers
[1]
.Groupingwithbias
Huttenlocher
[4]
,Nock
[2]
andNielsenandNock
[5]
make
isbasicallyonestepfurthertowardstheintegrationofthe
greedylocaldecisionstomergetheconnexcomponentsofuserintheloop,comparedtothemethodofgeneralexpec-
inducedsubgraphs.Sincesegmentationisaglobaloptimiza-tationsofagoodsegmentationintegratedinnon-purposive
tionprocess,theformerapproachisaprioriagoodcan-grouping
[9]
;itissolvedbypointingintheimagesomepix-
didatetotackletheproblem,evenwhenitfacescomputa-els(the
bias
)thattheuserfeelbelongtoidentical/different
tionalcomplexityissues
[3]
.However,strongglobalproper-
objects,andthensolvingthesegmentationasaconstrained
tiescanbeobtainedforthelatterapproaches,suchasqual-groupingproblem:pixelswithidenticallabels
must
belong
itativeboundsontheoverallsegmentationerror
[4,2]
,or
tothesameregioninthesegmentation’sresult,whilepixels
evenquantitativebounds
[5]
.
withdifferentlabels
mustnot
belongtothesameregion.The
7191213252729213
135791131517191123252729213335373931434457494153555
RAITPR2195
LCENIRPESR.Nock,F.Nielsen/PatternRecognition()–
solutionfortheglobalapproachofYuandShi
[1,7]
ismath-
ematicallyappealing,butitiscomputationallydemanding,
anditrequiresquiteanextensivebiasforgoodexperimental
resultsonsmallimages.Furthermore,itmakesitdifficultto
handletheconstraintsthatsomepixelsmustnotbelongto
thesameregions;thatiswhythistechniqueismainlyused
fortheparticularbiasedsegmentationprobleminwhichone
wantstosegregatesomeobjectsfromtheirbackground.
Inthispaper,weproposeageneralsolutiontobiased
grouping,basedonalocalapproachtoimagesegmentation
[2,5]
,whichbasicallyconsistsinusingthebiasfortheesti-
mationofnon-parametricmixturemodels.Distribution-free
processingtechniquesareuseful,ifnotnecessary,ingroup-
ing
[10,2]
.However,estimatingclustersindataisalready
farfrombeingtrivialevenwhensignificativedistribution
assumptionsareassumed
[11]
.Thisiswherethebiasisof
hugeinterestwhenitcomestogrouping,asbiasdefines
partiallylabeledregions,withtheconstraintthatdifferent
labelsbelongtodifferentregions.TheapproachofNock
[2]
andNielsenandNock
[5]
isalsoconceptuallyappealing
foranadaptationtobiasedsegmentation,becauseitconsid-
ersthattheobservedimageistheresultofthesamplingof
atheoreticalimage,inwhichregionsarestatisticalregions
characterizedbydistributions.Thereisnodistributionas-
sumptiononthestatisticalpixelsofthistheoreticalimage.
Theonlyassumptionmadeisanhomogeneityproperty,ac-
cordingtowhichtheexpectationofasinglecoloristhe
sameinsideastatisticalregion,anditisdifferentbetween
twoadjacentstatisticalregions.Thesegmentationproblem
isthustheproblemofrecognizingthepartitionofthesta-
tisticalregionsonthebasisoftheobservedimage.Thebi-
asedgroupingproblemturnsouttoallowstatisticalregions
tocontaindifferentstatisticalsub-regions,notnecessarily
connected,eachsatisfyingindependentlythehomogeneity
property,
and
forwhichtheuserfeelstheyallbelongtothe
sameperceptualobject.Thus,ityieldsasignificantgeneral-
izationofthetheoreticalframeworkofNock
[2]
andNielsen
andNock
[5]
.
Ourcontributioninthispaperistwofold.First,itconsists
oftwomodificationsandimprovementstotheunbiasedseg-
mentationalgorithmofNock
[2]
andNielsenandNock
[5]
.
Theiralgorithmcontainstwostages.Informally,itsfirstpart
isaprocedurewhichordersasetofpairsofadjacentpix-
els,accordingtotheincreasingvaluesofsomereal-valued
function
f
.Itssecondpartconsistsofasinglepassonthis
order,inwhichitteststhemergingoftheregionstowhich
thepixelsbelong,usingaso-calledmergingpredicate
P
.
f
and
P
arethecornerstonesoftheapproachofNock
[2]
and
NielsenandNock
[5]
.Weproposeinthispaperabetter
f
,
andanimproved
P
relyingonaslightlymoresophisticated
statisticalanalysis.Oursecondcontributionistheextension
ofthisalgorithmtogroupingwithbias.Ourextensionkeeps
bothfastprocessingandthetheoreticalboundsonthequal-
ityofthesegmentation.Experimentally,theresultsappear
tobeveryfavorablewhencomparingthemtothoseofthe
approachofYuandShi
[1,7]
.