彩色图像边缘检测Color Edge Detection in Presence of Gaussian Noise.docx
《彩色图像边缘检测Color Edge Detection in Presence of Gaussian Noise.docx》由会员分享,可在线阅读,更多相关《彩色图像边缘检测Color Edge Detection in Presence of Gaussian Noise.docx(41页珍藏版)》请在冰豆网上搜索。
彩色图像边缘检测ColorEdgeDetectioninPresenceofGaussianNoise
ColorEdgeDetectioninPresenceofGaussianNoise
UsingNonlinearPre-filtering
Abstract:
AnewwayforedgedetectionincolorimagesdisturbedbyGaussiannoiseis
presented.Thispaperproposedamethodwhichusesamulti-passprocessingmethod
thatreducesthenoiseinthecolorimage.Themainideaforthispaperisthatadopttwo
differentpre-filteringthataimatavoidingfalseedgesproducedbynoiseandat
preservingtheimagedetailsduringnoiseremoval.Afterdoingthatweneeddesignan
algorithmforedgedetectionwhichaimsatdecreasingthesensitivitytonoiseofthe
overallmethodsothatwecansuccessfullyaccurateedgemapsevenencounterhighly
corrupteddata.Accordingtothesimulationanddata,wecanfindtheproposedmethod
fornoiseimagesarequitegood.
KeyWord:
edgedetection,fuzzysystems,Gaussiannoise,imageprocessing,nonlinear
filters.
INTRODUCTION
Edgedetectionisameasurewhichplaysaveryimportantroleintherealizationofa
completeimageunderstandingsystem(Likeimagesegmentationandobjectrecognition
arealldependsonthequalityoftheedgedetection).Comparedwithgrayscalepicture,
colorimagehasrichermeasurementinformationthatcanimprovethequalityofthe
imageandextenditsrange.However,edgedetectionwillencountersometroublewhen
imagehassomenoiselikeGaussian.Therearesomemethodsfocuseonmonochrome
pictureslikeSobeltechniquetothethreeimagechannels.ActuallySobel’sedgepointin
thecolorimagecouldbeestimatedbyevaluatingthemaximumofthegradient
componentssumsothatitcanyieldanaccurateedgemap.SomemethodslikeDifference
vector(DV)methodusinggradient-basedalgorithmswhichisverysensitivetonoisehas
agoodperformancetotheedgedetection.Butthesemethodsarelimited,especiallywhen
theimagedataarehighlycorrupted,theresultwillbecomeveryannoyingsoweneed
newmethodtosolvethisproblem.Newmethodwhichusesmulti-passprocessingaimsat
reducingnoisebeforeextractingtheimageedges.Combiningpre-filteringschemewith
algorithmforedgedetectionwhichnotonlyimprovetheincreasetheaccuracyofthe
edgedetectionbutalsoconsiderthenoisesmoothingandspatialresolution.Sothe
followswillshowthemethodologyandtheresult.
Methodology
Pre-filteringforedgedetection
Firstweneedknowpre-filteringforedgedetectionisbasedonthedifferentkindsof
noisypixels:
TypeApixels:
pixelscorruptedbynoisewithamplitudenottoodifferentfromthatofthe
neighbors.
TypeBpixels:
pixelscorruptedbynoisewithamplitudemuchlargerthanthatofthe
neighbors.
Fig1Blockdiagramofthemulti-passprocessing
(p)
WedealwithdigitizedmultichannelRGBimages.xdenotetheMulti-channelimageat
thepass,andthexdenotestheinputnoisyimage,xkmeanswheretheR(red),G
(0)
(green),andB(blue)channelsarerespectively,denotedbyk=1,k=2,k=3.Putthem
togetherwhichconstructthisdiagram.Inthismethod,weuse24-bitcolorimage.Sowe
haveL=256.
Fig2colordivision
HereweneedmentionthathereweusetypeApre-filteringtwicewhichaimstoincrease
theeffectivenessofthesmoothingaction.
Theproposedmulti-passprocessinginvolvesthefollowingoperations:
TypeAPre-filtering:
Differencesbetweenthepixeltobedealwithanditsneighborsasbelow:
Smalldifferencesaresupposednoisetobereduced;largedifferencesaresupposededges
tobepreserved.
Theprocedureisdefinedbythefollowingrelationship:
1
8
1
8
1
1
∑∑
x
(1)
k
(i,j)=xk
(0)
(i,j)+
(i,j)+
ς
ς
(1)
(xk
k
(0)
(i+m,j+n),xk
(0)
(i,j))
(i,j))
m=-1n=-1
1
1
∑∑
x
(2)
k
(i,j)=xk
(1)
(2)
(xk
k
(1)
(i+m,j+n),xk
(1)
m=-1n=-1
Whereςkisaparameterizednonlinearfunction.Letk(p)(i,j)beaconstantpixel
(p)
x
value,0≤xk(i,j)≤L-1atlocation[i,j].
(p)
ThisformulameansweputthenoiseimageintothetypeApre-filteringandLetx(n)be
thepixelluminanceatlocationn=[n1,n2]andthepixelsprocessedcombinewithits
neighborconstruct3*3window.Withthismodelwhichshowsbelowwecanuse
parameterizednonlinearfunctiontoachievethesmoothingactionandexcludesthevalues
thatareverydifferentfromthecentralelement,inordertoavoiddetailblurduringnoise
cancellation.
Fig33*3window
Theς(p)
k
parameterizednonlinearfunctionprocedureshowsasfollow:
ς
k(p)(u,v)=u-v
|u-v|≤a
(p)
k
Whereurepresentstheneighborandvdenotedthecentralpixel.
Withthisvaluewhichmeanstheneighborarequiteclosetothecentralpixel.Sothefilter
performsthearithmeticmeanofthepixelluminanceintheneighborhood,thus
realizingthemaximumsmoothingaction.
ς
k(p)(u,v)=0
|u-v|≥3a
(p)
k
Inthisformwhichmeanstheneighborisquitedifferencewiththecentralpixelthefilter
behavesastheidentityfilter,thusperformingthemaximumdetailpreservation.
(u,v)=(3ak(p)
-u+v)sgm(u-v)
ς
(p)
k
ak(p)≤u-v|≤3ak(p)
|
2
Intermediatesituationsareprocessedasacompromisebetweentheseoppositeeffects.
(p)
Accordingtotheformula,wefindthatthebehaviordependsonthevalueofak.Inorder
togettheoptimalvalue,weusetheautomaticparametertuningtogetthebestvalue.
Fig4Automaticparametertuning
.
(p)
HereweusePinsteadofak,andthekisequaltotheinitialpvalue.Herepisfrom2to
L/4,eachimagehasitsownpvalue,usingthistuningprocedurewecangetthevalue.Why
wechooseMSE?
Becauseitisconceptuallysimpleandwidelyused,followthestepsevery
time,throughdifferentk1andk2wecangettheoptimalvalueanddothesecondtuning
procedure.Aftergettingthepvaluewecanuseittodothesmoothingaction.
ThisistheresultfortypeApre-filtering:
Fig.5GaussiannoiseimageandnoiseimagewithtypeApre-filtering
Accordingtotheimage,wecanfindthatthenoiseimagewithtypeApre-filtering
maketheimagemoresmoothingandleavetheoutlierfortypeBpre-filteringtodeal
with.
TypeBpre-filtering:
ActuallytypeBpre-filteringfocusontheoutlier,weknowoutlierpresentinthedataasan
effectofthetailofGaussiandistributiontheeffectcanbecomeveryannoyingTypeBpre-
filteringconsidersthedifferencesbetweenthepixeltobeprocessedanditsneighborsina
differentway:
ifallthesedifferencesareverylarge,thepixelisanoutliertobecancelled.It
isbrieflysummarizedasfollows.
µLA(u,v)denotesthemembershipfunctionthatdescribesthefuzzyrelation:
“uismuch
largerthanv”andtheformulashowsbelow:
Usingthesethreeformulascanachievethegoalwhichoutliercanberemovedbythisway.
Wecanseetheresultshowsbelow:
Fig6imageafterfilterAandimageafterfilterB
IfweaddpepperandsaltnoisetooriginalimageandseehowtypeBworks.
BecausepepperandsaltnoisebelongtotypeBnoise,typeBfiltershouldhaveaverygood
performance.
Fig7pepperandsaltfiltering
AccordingtotheFig6and7wecanfindthatusethetypeBpre-filteringwhichcanremove
theoutliersuccessfullybutactuallyitstillnotnoisefree.Howeverwecanstilldeclarethis
methodisquitegood.
Noise-protectededgedetector
Eventhoughthenonlinearpre-filteringsignificantlyreducestheGaussiannoise,according
totheimagewhichshowsbeforetheimagedataarenotnoise-free.Thus,anedgedetector
withverylowsensitivitytonoiseisnecessary.Weusethefollowrelationshiptodothe
detection:
y(i,j)=(L-1)[1-MIN{μSM(B1),μSM(B2)}]
Suitablechoicesare:
S1={x
(3)
(i-1,j-1),x
(i-1,j-1),x
(i,j),x(i,j+1),x(i+1,j),A1,A2}
(3)
(3)
(i-1,j),x(i-1,j+1)}
(3)
S2={x
(3)
(3)
(3)
(i,j-1),x(i+1,j-1)}
S1={x
(3)
(3)
μSM
isthemembershipfunctionoffuzzyset“small,”symbol||||denotestheEuclidean
distance,ShisthesetofNh
Whereb1andb2areintegerparameters.Smallervaluesperformastrongeractivationofthe
operatorinthepresenceofimageedgesatthepriceofanincreaseofthesensitivityto
noise,sothechoosingofthistwovaluesalsoneedtobeconsideredinourexperiment.
Typicallyb1=5,30Theedgedetectionprocedureaimsatrepresentingobjectcontoursasbrightlinesinthe
resultingedgemap.Conversely,uniformregionsarerepresentedasdarkareas.
Itisknownthataccuratedetectionofedgesinnoisydatashouldsatisfiedtworequirements:
(1)Theedgedetectionprocessshouldavoidfalseedgesproducedbynoise
(2)Ensurethatactualedgesarecorrectlydetected.
Wewilldosummationbygrayandcolorimagetoshowhowouredgedetectorworks
Wecanseetheresultshowsbelow:
Fig8Summationbygrayimage
Fig9Summationbycolorimage
Fig10Summationbylenaimage
AccordingtotheFig8,9and10wecanfindthatbyusingmultichannel,colorimagehasa
betterpre-filteringandedgedetectionresult,becausethatcolorimagehas3channelsbut
grayimageonlyhas1channel,sothattypeAfilterwillhaveabetterperformancewhen
dealwithcolorimage.Wecandeclareourmethodisquitegood.
Wewillalsoshowtheedgedetectionresultwhenimageidcorruptedbypepperandsalt
noise.Theresultshowsbelow:
Fig11EDofpepperandsaltnoisyimage
Accordingtothefigure11,wecangetaverygoodedgedetectionresultbyusingtypeB
filter,italmostremoveallsuperimposingimpulsenoise(saltandpeppernoise).Thethree
mainpartsofourmethod:
multichannel,typeAfilter,typeBf