98
S.W.Franklin,S.E.Rajan/AppliedSoftComputing22(2014)94–100
representedasn,andifLirepresentsthenumberofneuronsintheithlayer,thenthetotalnumberofweightsisgivenby:
1N
n=
LiLi+1
where,i=1,2,...,n
(1)
i=1
Thelearningcapacity(C)isanimportantparameterwhichdeterminestheabilityofasystemtolearn.Forathreelayernet-work,thelearningcapacityisboundedby:
nh
log(nh2)
<C<nhlog(h)
(2)
where,histhenumberofhiddennodesandnisthenumberofinputnodes.
Theneuralnetworkappliedisathreelayerperceptronhavingoneinputnode, vehiddenlayersandoneoutputnode.Beforepresentingtheinputtotheneuralnetwork,preprocessingoftheimagemustbedone.Theneuralnetworkisconsideredtobetrainedwhenthesystemerrorfallsbelowanacceptablethresholdlevelandatthisstage,thevesselsareidenti edfromthebackgroundandthussegmentedimagesareobtained.
Algorithm
Neuralnetworkusingbackpropagationalgorithmisapplied,using5/6thofthedatafortrainingand1/6thofthedataforvali-dation[29,30].Thetrainingalgorithmusedinthebackpropagationnetworkisasfollows:
Step1–Readtheinputimage.
Step2–SeparatetheRGBcomponentsinR,G,Bdifferentplanes.Step3–Morphologicalopeningoperationisperformedto llthevesseldiscontinuities.
Step4–Preprocessingstepslikemean lteringandGaussian lteringareperformed.
Step5–RGBfeaturesareextractedforthepreprocessedimageanditsmeaniscalculated.
Step6–Grayscaleimageisobtainedforfurtherprocessing.Step7–Gaborfeaturesareextractedatdifferentorientations.Step8–Momentinvariants-basedfeaturesareextracted.
Step9–Samplesaretakenfromvesselandnonvesselregionsandtheyaretrainedusingneuralnetworks.
Step10–Segmentationofbloodvesselsisobtainedasoutput.Foreachtrainingpair,theneuralnetworkperformsthefollow-ingtasks.Atthe rststage,duringtheinitializationofweights,afewsmallrandomvariablesareassigned.Inthenextstage,i.e.,feedforwardstage,eachinputunit(Xi)receivesaninputsignalandtransmitsthissignaltoalltheunitsinthelayerabovei.e.hiddenunits,z1,...,zp.Eachhiddenunitcalculatestheactivationfunction,
Z nj=f(zinj)where,z inj=Voj+
XiVij
(3)
i=1
Vojisthebiasonhiddenunitj.Eachhiddenunitsendsthissignaltoalloutputunits.
Eachoutputunit(yk,k=1...m)thensumsitsweightedinputsignalsandappliesitsactivationfunctionYktocalculatetheoutputsignals.
Y pk=f(y ink)where,y ink=wok+
zjwjk
(4)
j=1
wokisthebiasonoutputunitk.
Duringbackpropagationoferrors,eachoutputunitcomparesitscalculatedactivationwithitstargetvaluetk,to ndouttherelatederrorsforthatpattern.
Errorinformationtermisgivenas,
k=(tk yk)f (y ink)
(5)
Itisusedtodistributetheerrorattheoutputunitykbacktoallunitsinthepreviouslayer.Theneachhiddenunit(zj,j=1,...,n)sumsitsdeltainputsfromunitsintheabovelayeras:
m inj=
jwjk
(6)
k=1
Similarlytheerrorinformationterm j= injf(z inj)iscom-putedforeachhiddenunit.Duringthe nalstage,theweightsandbiasesareupdatedusingthe factorandtheactivationfunctionisupdated,using asthelearningrate.Eachoutputunitupdatesitsweightsandbiasesbyusingtheweightcorrectionterm Wjkandbiascorrectionterm Wok,
where, Wjk= kzjand Wok= k
(7)Therefore,Wjk(new)=Wjk(old)+ Wjkand
(8)Wok(new)=Wok(old)+ Wok
(9)
Similarlytheweightsandbiasesareupdatedforeachhiddenunit.Finally,thestoppingconditionistested.Thestoppingcondi-tionmaybeminimizationoftheerrors,numberofepochsetc.
Experimentalresultsanalysis
TheautomaticvesselsegmentationalgorithmproposedinthispaperwastestedwiththeimagesofDRIVEdatabase.Thistech-niquehasbeenprovedasoneofthemostvaluabletoolsforthesegmentationofbloodvesselsinretinalimages.Illustrativeves-selsegmentationresultsforafewimagesfromDRIVEdatabaseareshowninFig.5.
Theneuralnetworkbasedvesselsegmentationusingbackprop-agationalgorithmemployedinthispaperhasacomputationallydemandingtrainingphase.Nevertheless,itguaranteesanef cientvesselclassi cationphaseandhasprovenexcellentperformance.Inaddition,thetechniquespresentedhereareconceptuallyquiteeasierandcanbeimplementedveryeffectively.Ourtechniquecomesunderthebasicclassi cationofpixelprocessing-basedapproach.Theoretically,onlytheproperselectionofinitialweightswouldresultinsuccessfulnetworktrainingandensurebetterclas-si cationperformance.Itisfoundthattheperformanceofthenetworkimproveswithincreasednumberofrandomlyselectedtrainingsamples.Basedontheexperimentalresults,itisseenthatifthenumberofhiddenlayersincreases,theconvergencebecomesslowerandthetrainingbecomestedious.Itisalsoobservedthat,ifthenumberofhiddenlayerincreases,alargerlearningrateisessential.However,ifthelearningratebecomestoolarge,itresultsinoscillations.
Performanceevaluation
WehavetestedandevaluatedourmethodusingtheimagesfromDRIVEdatabase.Forthispurpose,thetrainingsetwasformedbypixelsamplesfromthetwentylabeledtrainingimages.TestswereperformedbyapplyingGabor ltersandvaryingtheGabororientationsatdifferentangles.Predictionofthemomentinvariants-basedfeatureswasalsoconsidered.Finally,testswereperformedbytakingsamplesfromvesselandnonvesselregionsandthesedataweretrainedusingneuralnetworks.Inourmethod,wehaveselectedsegmentationaccuracyastheperformancemea-sure.TheseperformancemeasureswereevaluatedusingROCcurves,whichareplotsoftruepositivefractionsversusfalseposi-tivefractionsforvaryingthresholdsontheposteriorprobabilities.
ICPR06_A new approach to automated retinal vessel segmentation using ...Gabor filters and scale production, and a threshold probing technique ...
热门文档