原创版图像增强外文文献及翻译.docx

原创版图像增强外文文献及翻译.docx

《原创版图像增强外文文献及翻译.docx》由会员分享，可在线阅读，更多相关《原创版图像增强外文文献及翻译.docx（15页珍藏版）》请在冰豆网上搜索。

原创版图像增强外文文献及翻译

附录A：

外文文献

AnEffectiveAutomaticImageEnhancementMethod

ABSTRACTOtsumethodispropertodealwithtwoconditions:

（1）twoormoreclasseswithdistintivegray-valuesrespectively;

（2）classeswithoutdistinctivegray-values,butwithsimilarareas.However,whenthegray-valuedifferencesamongclassesarenotsodistinct,andtheobjectissmallrelativetobackgroud,theseparabilitiesamongclassesareinsufficient.Inordertoovercometheaboveproblem,thispaperpresentsanimprovedspatiallow-passfilterwithaparameterandpresentsanunsupervisedmethodofautomaticparameterselectionforimageenhancementbasedonOtsumethod.Thismethodcombinesimageenhancementwithimagesegmentationasoneprocedurethroughadiscriminantcriterion.Theoptimalparameterofthefilterisselectedbythediscriminantcriteriongiventomaximizetheseparabilitybetweenobjectandbackground.Theoptimalthresholdforimagesegmentationiscomputedsimultaneously.Themethodisusedtodetectthesurfacedefectofcontainer.Experimentsillustratethevalidityofthemethod.

KEYWORDSimageprocessing;automatedimageenhancement;imagesegmentation;automatedvisualinspection

1Introduction

Automatedvisualinspectionofcrackedcontainer（AVICC）isapracticalapplicationofmachinevisiontechnology.Torealizeourgoal,fouressentialoperationsmustbedealtwith–imagepreprocessing,objectdetection,featuredescriptionandfinalcrackedobjectclassification.Imageenhancementistoprovidearesultmoresuitablethanoriginalimageforspecificapplications.Inthispapertheobjectiveofenhancement,followedbyimagesegmentation,istoobtainanimagewithahighercontentabouttheobjectinterestingwithlesscontentaboutnoiseandbackground.Gonzalez[1]discussesthatimageenhancementapproachesfallintotwomaincategories,inthatspatialdomainandfrequencydomainmethods.Burton[2]appliesimageaveragingtechniquetofacerecognitionsystem,makingitabletorecognisefamiliarfaceseasilyacrosslargevariationsinimagequality.Centeno[3]proposesanadaptiveimageenhancementalgorithm,whichreversetheprocessingorderofimageenhancementandsegmentationinordertoavoidsharpeningnoiseandblurringborders.Munteanu[4]appliesartificialintelligencetechnologytoimageenhancementprovidingdenoisingfunction.Inadditiontospatialdomainmethods,frequencydomainprocessingtechniquesarebasedonmodifyingtheFouriertransformofanimage.Bakir[5]discussesimageenhancementusedformedicalimageprocessinginfrequencyspace.Besides,Wang[6]presentsaglobalmultiscaleanalysisofimagesbasedonHaarwavelettechniqueforimagedenoising.Recently,Agaian[7]proposesimageenhancementmethodsbasedonthepropertiesofthelogarithmictransformdomainhistogramandhistogramequalization.Weapplyspatialprocessinghereinordertoguaranteethereal-timeandsufficientaccuracypropertyofthesystem.

Segmentationisdiscussedin[8].Themostsimplest,representedbyOtsu[9],ismethodusingonlythegraylevelhistogramanalysistomaximizetheseparabilityoftheresultantclasses.Kuntimad[10]describesamethodforsegmentingdigitalimagesusingpulsecoupledneuralnetworks（PCNN）.Salzenstein[11]dealswithacomparisonofrecentstatisticalmodelsonfuzzyMarkovrandomfieldsandchainsformultispectralimagesegmentation.Duetoill-defined,thereisnouniquesegmentationofanimage.Evaluationofsegmentationalgorithmsthusfarhasbeenlargelysubjective.Ranjith[12]demonstrateshowarecentlyproposedmeasureofsimilaritycanbeusedtoperformaquantitativecomparisonamongimagesegmentationalgorithms.

Inthispaper,wepresentanimprovedspatiallow-passfilterwithatunableparameterinthemaskmakingallelementsnolongersumtounity.Theoptimalparameterforthefiltercanbedeterminedbytheimproveddiscriminantcriterionbasedontheonementionedin[9].Convolvingimageswiththismask,thebackgrounduninterestingcanberemovedeasilyleavingtheobjectintacttosomeextent.Theremainderofthepaperisorganizedasfollows:

Sect.2presentshowtoenhanceaninputimageintheoryandpresentsthealgorithm.Sect.3illustratesthevalidityofthemethodinSect.2.Finally,conclusionanddiscussionarepresentedinSect.4.

2　ImageEnhancement

2.1AnalysisofPriorKnowledge

Thepreprocessingqualityinfluencesthelatterworkdirectly,inthat,featuredescription.Therefore,analysisforthecharacteristicsrelatedtoinputimagesshouldbepresented.AstandardimageofcrackedcontainerisshownasFig.1（a）.Fromtheimage,weseethecrackedpartoccupiessmallregion.Muchnoise,suchasrust,shadow,smearetc,appearswithinthebackground.Atacoarseglance,however,wefindgrayleveloftheholeislessthantheotherpartsdistinctly.Furtherstudyshowsgraylevelofpixels,aroundtheedgeofthehole,istheminimal.Fig.1（b）displaysthehistogramofFig.1（a）andedgeoftheholeismarked.

Fig.1（a）isastandardgraylevelimageofacrackedcontainer（b）isthehistogramofFig.1（a）,indicatinggraylevelregionofthehole’sedge.

2.2Formulation

Thissectiondiscussestheprincipalcontentinthepaper.Traditionalspatialfilterusesa3×3mask,theelementsofwhichsumtounity,toconvolvewiththeinputimage.Thismethodcandealwithsomecasesshowninequation

（1）:

（1）

where,Iisimageinterested,NisGaussianwhitenoise,（x,y）denoteseachpairofcoordinates.NcanbedeliminatedbyblurringG.Ourobjective,however,istodeliminatenotonlywhitenoise,butanyotherbackgrounduninteresting.Thusequation

（1）isimprovedbyequation

（2）:

（2）

where,I'istheobject,N'consistsofwhitenoiseandtheotherpartsexceptI'.Fig.2（c）displaysanimprovedmaskwithaparameterPara.WewilllaterillustratethattuningParaproperlyistofacilitateobjectsegmentation.Thesmoothingfunctionusedisshowninequation（3）:

（3）

where,F（x,y）denotesthesmoothingfilter,inthat,themaskshownasFig.2（c）.

Now,weonlyconsidergray-levelimages,anddefineMgasthemaximumgraylevelofanimage.Thenthefollowingequationsaresettodistinguishtheobjectofinterestandthenon-object:

（4）

Inessence,convolutionoperatorisalow-passfilteringprocess,whichblursanimagebyslidingamaskthroughtheimageandleavesthefilteringresponseatthepositioncorrespondingtocentrallocationofthemask.Onequestionoccursthat,whynotenhancevalueofeachpixelbythesamescaledirectlyforthedistinctgraylevelsbetweentheobjectandbackground.Thereasonisthatitdoesn’tconsidertherelationshipofadjacentpixels.Whenindividualnoisepointoccur,enhancingitsgrayvaluedirectlywillpreservethenoisepoint.Experimentsillustratethelattermethodwillleavelotsofnoisepointscan’tberemoved,buttheformermethodwillnot.

Now,wewillsearchtheoptimalparameterParasoastomaximizetheseparabilitybetweenobjectandbackground.LetagivenimageberepresentedinLgraylevels.ThenumberofpixelsatleveliisdenotedbyniandthetotalnumberofpixelsbyN.TheprobabilityofeachlevelisdenotedbyPiasfollow[9]:

（5）

SupposethatwepartitionthepixelsintotwoclassesC0andC1（objectandbackground）byathresholdatlevelk;C0denotespixelswithlevels[1,…,k],andC1denotespixelswithlevels[k+1,…,L].Thentheprobabilitiesofclassoccurrencew0,w1andtheclassmeanlevelsu0,u1respectively,aregivenby

（6）

（7）

（8）

（9）

（10）

（11）

（12）

Theprocedureofobtainingoptimalparaisbasedonobtainingoptimalthresholdforeveryfilteredimage.Theoptimalthresholdisdeterminedbymaximizingtheseparabilitybetweenobjectandbackgroundusingthefollowingdiscriminantcriterionmeasureasmentionedin[9]:

（13）

where

（14）

and

arethebetweenclassvarianceandthetotalvarianceoflevels,respectively.

（15）

Theoptimalthresholdk*thatmaximizesnisselectedinthefollowingsequentialsearchbyusingequation（5）-（14）:

（16）

Equation（16）isadiscriminantcriteriontoselectthegrayleveltomaximizetheseparabilitybetweenobjectandbackgroundforagivenpicture.Inthispaper,aparameterParaisintroduced,sotheequations（6）~（9）,（11）~（14）,（16）isparameterizedbyParaandkandequations（10）,（15）isparameterizedbyPara.Equation（13）canberewrittenas:

（17）

Where

isnotaconstantanymoreandisnotnegligible,butsomecomputationreductioncanbeoperatedon

and

Here,whatwewanttoacquireistheproperfilteredpictureincludingvividobjectbysearchingparameterPara,thediscriminantcriterionusedisimprovedasfollow:

（18）

Intheaboverepresentation,parameterParaplaysanimportantrole,becauseoptimalParamakestheseparabilitybetweenobjectandbackgroundmaximal,andmakeOtsusegmentationmethodeffectivetosegmentsmallobjectfromlargebackgroundwithoutdistinctivegray-valuebetweenthem,whichcanbeobservedlaterfromimagehistogramafterimageenhancement

2.3ExistenceDiscussionofParaandk*

Theproblemaboveisreducedtosearchforathresholdk*undertheconditionofParawhichmaximizesthediscriminantcriterioninequation（18）.Theconditiondiscussedistheimagewithtwoclassatleast.Subsequently,thefollowingtwocasesdon’toccur,inthat,

（1）w0orw1iszerooriginallywithoutsettingPara,inwhichthereisonlyoneclass;

（2）w0orw1iszerowithcertainincreasingPara,inwhichthereisalsooneclassfinally;

Theabovetwocasesaredecribedas:

ThecaseconcernedisA,Thus,thereiscertainParawithproperktomakediscriminantcriterionmaximal.

3Experiments

Thispaperaimsa