图像匹配几种常见算法外文文献.docx
- 文档编号:23963020
- 上传时间:2023-05-22
- 格式:DOCX
- 页数:36
- 大小:120.35KB
图像匹配几种常见算法外文文献.docx
《图像匹配几种常见算法外文文献.docx》由会员分享,可在线阅读,更多相关《图像匹配几种常见算法外文文献.docx(36页珍藏版)》请在冰豆网上搜索。
图像匹配几种常见算法外文文献
SystemsDesignEngineering,UniversityofWaterloo,200UniversityAvenueWest,Waterloo,Ontario,CanadaN2L3G1
Abstract
Aninterestingprobleminpatternrecognitionisthatof
image
registration,whichplaysanimportantroleinmanyvision-basedrecognitionandmotionanalysisapplications.Ofparticularinterestamongregistrationproblemsaremultimodalregistrationproblems,wherethe
images
existindifferentfeaturespaces.State-of-the-artphased-basedapproachestomultimodal
image
registrationmethodshaveprovidedgoodaccuracybuthavehighcomputationalcost.Thispaperpresentsa
fast
phase-basedapproachtoregisteringmultimodal
images
forthepurposeofinitialcoarse-grainedregistration.Thisisaccomplishedbysimultaneouslyperformingbothgloballyexhaustivedynamicphasesub-cloud
matching
andpolynomialfeaturespacetransformationestimationinthefrequencydomainusingthe
fast
Fouriertransform(FFT).Amultiscalephase-basedfeatureextractionmethodisproposedthatdeterminesboththelocationandsizeofthedynamicsub-cloudsbeingextracted.Asimpleoutlierpruningbasedonresamplingisusedtoremovefalsekeypointmatches.Theproposedphase-basedapproachtoregistrationcanbeperformedveryefficientlywithouttheneedforinitialestimatesorequivalentkeypointsfromboth
images.
Experimentalresultsshowthattheproposedmethodcanprovideaccuraciescomparabletothestate-of-the-artphase-based
image
registrationmethodsforthepurposeofinitialcoarse-grainedregistrationwhilebeingmuch
faster
tocompute.
Keywords:
Image
registration;Phase;
Fast
Fouriertransform;Multimodal;Keypoints;Dynamicsub-clouds
ArticleOutline
1.Introduction
2.Multimodalregistrationproblem
3.Previouswork
4.Proposedregistrationalgorithm
4.1.Keypointdetectionandsub-cloudsizeestimation
4.2.Phasesub-cloudextraction
4.3.Simultaneoussub-cloudmatchingandfeaturespacetransformationestimation
4.4.Solvingthesimultaneousmatchingandfeaturespacetransformationestimationprobleminthefrequencydomain
4.5.Outlierpruningthroughresampling
4.6.Algorithmoutline
5.Computationalcomplexityanalysis
6.Experimentalresults
7.Conclusionsandfuturework
Acknowledgements
References
1.Introduction
Imageregistrationistheprocessofmatchingpointsinoneimagetotheircorrespondingpointsinanotherimage.Theproblemofimageregistrationplaysaveryimportantroleinmanyvisualandobjectrecognitionandmotionanalysisapplications.Someoftheseapplicationsincludevisualmotionestimation[1]and[2],vision-basedcontent-basedretrieval[3]and[4],imageregistration[5],[6],[7]and[9],andbiometricauthentication[10].Inthebestcasescenario,theimagesexistatthesamescale,inthesameorientation,aswellasrepresentedinthesamefeaturespace.However,thisisnotthecaseinmostreal-worldapplications.Therearemanysituationswheretheimagesexistindifferentfeaturespaces.Thisparticularproblemwillbereferredtoasthemultimodalregistrationproblemandisaparticularlydifficultproblemtosolve.Examplesofthisprobleminreal-worldsituationsincludemedicalimageregistrationandtrackingofMRI/CT/PETdata[11]andbuildingmodelingandvisualizationusingLIDARandopticaldata[12]and[13].
Thereareseveralimportantissuesthatmakemultimodalregistrationadifficultproblemtosolve.First,manyregistrationalgorithmsrequirethatequivalentkeypointsbeidentifiedwithineachimage.However,giventhedifferencesbetweenfeaturespacesinwhichtheimagesexist,itisoftenaverydifficulttask.Thesignificantdifferencesbetweenfeaturespacesalsomakeitimpracticaltoperformdirectintensitymatchingbetweenthetwoimages.Inrecentyears,aneffectiveapproachtomultimodalregistrationhasbeenproposedthatutilizeslocalphase[14]and[15].Thisstate-of-the-artapproachevaluatesthemutualinformationbetweenthelocalphaseoftwoimagestodeterminetheoptimalalignmentandhasbeenshowntobeveryeffectiveatmatchingmultimodalmedicalimagedata,outperformingexistingmultimodalregistrationmethods[14]and[15].However,thisapproachiscomputationallyexpensive(O(N6)forthemutualinformationevaluationprocess).Assuch,aregistrationmethodthatisabletotakeadvantageoflocalphaseinformationtodeterminepointcorrespondencesbetweenimageswhilebeingcomputationallyefficientishighlydesiredforthepurposeofinitialcoarse-grainedregistration.
Themaincontributionofthispaperisfastphase-basedregistrationalgorithmforaligningmultimodalimages.Theproposedmethodisdesignedtoprovideafastalternativetothephase-basedregistrationalgorithmproposedbyMelloretal.[15].Itisimportanttonotethatthemaincontributionsofthispaperresideinthemethodsforkeypointdetectionanddynamicsub-cloudextraction,aswellasthemethodforsimultaneousphasecorrespondenceevaluationandfeaturetransformationestimation,notintheoutlierpruningscheme.Furthermore,theproposedmethodisdesignedforfastinitialcoarse-grainedmatchingandbynomeansguaranteethesmoothnessoftheglobaldatacorrespondenceproblem.Afine-grainedmatchingmethodcanbeusedaftertheinitialmatchingtoprovideimprovedalignmentbasedonglobalsmoothnessconstraints.
2.Multimodalregistrationproblem
Themultimodalregistrationproblemcanbedefinedinthefollowingmanner.Supposethereexisttwoimagesfandg,wherepointsinfandgarerepresentedusingtwodifferentfeaturespaces,respectively.Foreverypointinf,thegoalofregistrationistodetermineacorrespondingpointingsuchthatthehighestdegreeofcorrespondencecanbefoundbetweenfandg.ThisproblemcanbealternativelybeformulatedasfindingtheoptimaltransformationTthatmapsallpointsfromftothepointsfromgsuchthatthehighestdegreeofcorrespondencecanbeachieved.Therelationshipbetweenfandgcanbedefinedas
(1)
where
and
arecoordinatevectorscorrespondingtofandg,respectively,andTisatransformationthatmapspointsfromftog.
Basedontheaboverelationship,themultimodalregistrationproblemcanbeformulatedasaminimumdistanceoptimizationproblem,withthedistancerepresentingthedegreeofdatacorrespondencebetweentwoimagesexpressedas
(2)
whereDisthedistancefunctionthatisinverselyproportionaltothedegreeofcorrespondencebetweenfeaturepoints.LowvaluesofDindicateahighlevelofcorrespondencebetweentheimages.
3.Previouswork
Alargenumberofmethodshavebeenproposedforthepurposeofimageregistration.Ingeneral,currentmethodscanbegroupedintofourmaintypes:
(1)Methodsbasedonrelativedistances[16],[17],[18]and[19].
(2)Methodsinthefrequencydomain[20],[21]and[22].
(3)Methodsbasedondirectcomparisons[6],[7],[8],[9],[23],[24],[25],[26]and[27].
(4)Methodsbasedonextractedfeatures[14],[15],[28],[29],[30],[31],[32],[33],[34],[35]and[36].
Methodsbasedonrelativedistancesexploitthespatialrelationshipsbetweenneighboringpixelswithinanimagetodeterminethebestmatchbetweentwopoints.Thesealgorithmsarebasedontheassumptionthatifapointinimagef,pf,0,hasacorrespondingpointinimageg,pg,0,thenthereexistotherpointsinf,{pf,1,pf,2,…,pf,n},thathaveacorrespondingpointsing,{pg,1,vg,2,…,pg,n},suchthatthedistancebetweenpf,0andpointpf,kisequaltothedistancebetweenpg,0andpointpg,k.Methodsbasedonrelativedistancesareprimarilyusefulforsituationswherethetransformationbetweentheimagesconsistsonlyoftranslationsandrotations.
Methodsinthefrequencydomain[20],[21]and[22]exploitthefrequencycharacteristicssuchasphasetoestimatethetransformationbetweentwoimages.Acommonfrequencydomainregistrationmethodisphasecorrelation,wheretheFouriercoefficientscalculatedfromimagefaredividedbythatcalculatedfromimageg.Performingtheinversetransformontheresultyieldsasinglepeakindicatingthetranslationthatmatchesthetwoimages.ThistechniquehasbeenextendedtoaccountforglobalrotationsandscalebyReddyetal.[21].Assuch,frequencydomainmethodsareonlysuitedforgloballyrigidpointcorrespondences.
Methodsbasedondirectcomparisons[6],[7],[8],[9],[23],[24],[25],[26]and[27]attempttofinddatapointcorrespondencesbetweentwoimagesbyperformingpointmatchesdirectlyintheirrespectivefeaturespaces.Manytechniquesinthisgroupmakeuseoffeatureinformationfromneighboringdatapointstodeterminethesimilaritybetweentwodatapoints.Somecommonsimilaritymetricsusedindirectcomparisonmethodsincludemaximumlikelihood[5],correlation[23]and[26],andmutualinformation[6],[7],[8]and[9].Ofparticularinterestinrecentyearsaretechniquesbasedonmutualinformation,whichattempttomatchdatapointsbyfindingthemutualdependencebetweentheimages.Thekeyadvantageoftechniquesbasedonmutualinformationisthatitallowsimagesexistingindifferentfeaturespacestobecomparedinadirectmanner.However,suchtechniquesalsorequiregoodinitialmatchestimatestoproduceaccurateresultsastheyareveryunder-constrainedinnature.Furthermore,techniquesbasedonmutualinformationarecomputationallyexpensiveandmaynotbepracticalforcertainsituationswherecomputationalspeedisimportant.Acomparisonbetweenthecomputationalcomplexityofmutualinformation-basedtechniquesandtheproposedmethodwillbediscussedlateroninthepaper.
Inmethodsbasedonextractedfeatures[14],[15],[28],[29],[30],[31],[32],[33],
- 配套讲稿:
如PPT文件的首页显示word图标,表示该PPT已包含配套word讲稿。双击word图标可打开word文档。
- 特殊限制:
部分文档作品中含有的国旗、国徽等图片,仅作为作品整体效果示例展示,禁止商用。设计者仅对作品中独创性部分享有著作权。
- 关 键 词:
- 图像 匹配 常见 算法 外文 文献