resolvedpatcha

BLOCK-ADAPTIVEINTERPOLATIONFILTERFORSUB-PIXELMOTIONCOMPENSATIONJaehyunCho,Dong-BokLee,ShinCheolJeong,ByungCheolSongSchoolofElectronicEngineering,InhaUniversityYonghyun-dong,Nam-gu,402-751,Incheon,RepublicofKoreaphone:+(82)32-860-7413,email:bcsong@inha.
ac.
krABSTRACTAdaptiveinterpolationfilteringforsub-pelmotionestima-tionisoneofseveralsuperiortechniquesofITU-TKTACODECtotheH.
264/AVCCODEC.
However,theadaptiveinterpolationfilteringhasalimitationincodingefficiencybecauseofitsframe-basedupdatestrategyoffiltercoeffi-cients.
Inordertoovercomesuchaproblem,thispaperpresentsablock-adaptiveinterpolationfilteringusinglearn-ing-basedsuper-resolution.
Theproposedblock-adaptiveinterpolationfilteringforquarter-pelmotionestimationconsistsoftwosteps:two-timesup-scalingofhalf-pelaccu-racyandsubsequenttwo-timesup-scalingofquarter-pelaccuracy.
Thedictionaryoptimizedforeachstepisem-ployedtoproducethepreciseup-scaledblocks.
Simulationresultsshowthattheproposedalgorithmimprovescodingefficiencyupto5.
3%incomparisonwiththepreviousadap-tiveinterpolationfilteringforKTA.
1.
INTRODUCTIONRecently,withrapiddevelopmentofsemiconductoranddigi-taldisplay,highdefinition(HD)videocontentshavebeenpopular.
Toefficientlytransmitorstoresuchhugevideodata,highcompressiontechnologyisrequired.
Forexample,H.
264/AVC[1]isthelatestvideocodingstandardtomeetsuchrequirement,whichwasjointlyimplementedbyITU-T(InternationalTele-communicationUnion)andMPEG(Mov-ingPictureExpertGroup).
AsakeycompressiontoolforH.
264/AVC,sub-pelmotioncompensationiscomposedofhalf-pelmotioncompensationusing6-tapfilteroffixedcoef-ficients,andsubsequentquarter-pelmotioncompensationusingbilinearinterpolation.
However,thefixedfiltercoeffi-cientsmayoftendeterioratecodingefficiencybecausetheynevertakeintoaccountspatialcharacteristicsofeveryframe.
Forarecentfewyears,VCEG(VideoCodingExpertsGroup)ofITU-ThasdevelopedKTA(KeyTechnologyArea)softwareasaninterimprocessfornext-generationvideocod-ingstandardbyevaluatingalotofnewcompressiontoolsandadoptinghighperformancetoolsamongthem.
AsoneofdominanttechniquesforKTA,AIF(AdaptiveInterpolationFilter)forsub-pelmotioncompensation[2-4]improvedcod-ingefficiencyuptoabout10%incomparisonwithitscoun-terpartofH.
264/AVC.
However,theAIFdoesnotsufficient-lyconsiderlocalcharacteristicsinaframebecauseitupdatesfiltercoefficientsonaframebasis.
ThisisamajordrawbackoftheAIFmethod.
Recently,MPEGhaslaunchednewcod-ingstandardtoreplaceH.
264/AVCinthefuture,whichiscalledHEVC(HighEfficiencyVideoCoding)[5].
Forsub-pelmotioncompensation,theup-to-dateversionofHEVCadaptivelychoosesonebetweenasimpleDCT(DiscreteCosineTransform)-basedinterpolationfilteranddirectionalinterpolationfilter.
TheinterpolationmethodforHEVCismeaningfulintermsofcomputationalcomplexity,butitdoesnotshowbettercodingefficiencythantheAIFforKTA.
Ontheotherhand,so-calledsuper-resolution(SR)algo-rithms[6-9]havebeendevelopedasthemostpromisingup-scalingapproach.
AtypicalSRmakesuseofsignalprocessingtechniquestoobtainahighresolution(HR)image(orasequence)frommultiplelow-resolution(LR)images.
Ingeneral,successofsuchSRschemesdependsonexistenceofsub-pixelmotionbetweenadjacentLRimagesandaccuratesub-pixelestimation.
However,sub-pixelmotionestimationamongneighborLRimagesrequiresnotonlyhugecomputa-tionalcost,butalsoitsaccuracyisnotguaranteedincertainenvironments.
Inordertosolvetheabove-mentionedprob-lem,alotofsingleimage-basedSRmethodssuchaslearn-ing-basedSRalgorithmshavebeendevised[7-9].
Ingeneral,learning-basedSRiscomposedoftwophases:Off-linelearn-ingphaseandon-linesynthesisphase.
Atthelearningphase,thetrainingdata,i.
e.
,dictionaryconsistingofLRandHRpatchesisconstructed.
TheLRandHRpatchpairsareob-tainedfromvarioustrainingimages.
Duringthesynthesisphase,theinputLRimageissuper-resolvedbyusingthedictionary.
ForeachLRpatchintheinputimage,itsnearestneighborLRpatchesareexploredfromthedictionary.
ThehighfrequencycomponentsoftheinputLRpatcharesynthe-sizedusingthebestmatchedLRpatches[9].
Sincethislearn-ing-basedSRprovidessuperiorvisualqualitytoconventionalFIRfiltersattheexpenseoflargememorysize,itcanbeanattractivesolutiontohighperformanceinterpolation.
Inordertoovercometheabove-mentioneddrawbackofthepreviousAIFmethod,thispaperproposesablock-adaptiveinterpolationfilter(BAIF)usinglearning-basedSR.
Theproposedalgorithmimprovestheperformanceofsub-pelmotioncompensationbyadaptivelyupdatingfiltercoeffi-cientsonablockbasiswithoutadditionalsideinformation.
TheBAIFconsistsoftwostepsforhalf-pelinterpolationandquarter-pelinterpolation.
Inoff-linelearningphase,theop-timaldictionaryofeachstepisderivedfromvariousLRandHRtrainingimages.
Simulationresultsshowthatthepro-19thEuropeanSignalProcessingConference(EUSIPCO2011)Barcelona,Spain,August29-September2,2011EURASIP,2011-ISSN2076-14652156posedalgorithmprovideshighercodingefficiencyofupto5.
3%thanthepreviousAIFforKTA.
2.
PREVIOUSWORKSThissubsectiondescribesseveralAIFmethodsforKTA.
NSAIF(Non-SeparableAIF)[2]interpolatesasubtwo-dimensional(2D)filtercoefficientsoptimizedatthepixelposition.
InFig.
1,aone-dimensional(1D)6isappliedtosub-pelssuchasa,b,c,d,h,andC1-C6areusedforthesub-pelpositionsa,b,cford,h,l.
Foreachoftheremainingsub-pelpositionsi,j,k,m,n,ando,the6ⅹ6filtercoefficientsarecalculated.
Forallsub-pelpositions,theoptimalfiltercoefficientsarecalculatedinawaythatthepredictionerrorenergyisminmized.
Thefiltercoefficientscanbeupdatedonaframebasis.
SinceSAIF(SeparableAIF)[3]isbasedontwo1Dfiltercoefficients,itcanachievelightinterpolationcomplexitywithoutanypenaltyoncodingefficiencyincomparisonwithNSAIF.
Notethatthecomputationalexpenseoftheitselfisreducedby24%incaseof4ⅹ4motionblocks.
AsanotherlowcomplexityAIF,DIF(DirectionalAIF)[employsasingle1Ddirectionalinterpolationfiltercoeffcientsateachsub-pellocation.
Thedirectionoftheinterpoltionfilterisdeterminedaccordingtothealignmentofthecorrespondingsub-pixelwithintegerpixelsamples.
Foreample,twosub-pelseandoofFig.
1areinterpolatedbyaplyinga6-tapfiltertosixintegersamplesA1,B2,C3,D4,E5andF6.
Sinceallsub-pelsareobtainedusingonly1Dfilteroperations,thecomplexityoftheDIFissignificantlylessthanitscounterparts.
Intheworstcase,theinterpolationcomplexityoftheDIFis1/3ofNSAIFandlessthan1/2ofSAIF.
Fig.
1.
Integersamples(shadedblockswithupperfractionalsamplepositions(whiteblockswithlowerposedalgorithmprovideshighercodingefficiencyofuptoPREVIOUSWORKSThissubsectiondescribesseveralAIFmethodsforKTA.
]interpolatesasub-pelwithltercoefficientsoptimizedatthedimensional(1D)6-tapfilterandl.
Thesamplesa,b,c,andA3-F3pelpositionse,f,g,6filtercoefficientsarecalculated.
pelpositions,theoptimalfiltercoefficientsarecalculatedinawaythatthepredictionerrorenergyismini-beupdatedonaframebasis.
]isbasedontwo1Dfiltercoefficients,itcanachievelightinterpolationcomplexitywithoutanypenaltyoncodingefficiencyincomparisonwithNSAIF.
NotethatthecomputationalexpenseoftheSAIF4motion-compensatedAsanotherlowcomplexityAIF,DIF(DirectionalAIF)[4]employsasingle1Ddirectionalinterpolationfiltercoeffi-pellocation.
Thedirectionoftheinterpola-filterisdeterminedaccordingtothealignmentofthepixelwithintegerpixelsamples.
Forex-ofFig.
1areinterpolatedbyap-A1,B2,C3,D4,E5,pelsareobtainedusingonly1Dfilteroperations,thecomplexityoftheDIFissignificantlylessthanitscounterparts.
Intheworstcase,theinterpolationcomplexityoftheDIFis1/3ofNSAIFandlessthan1/2of(shadedblockswithupper-caseletters)andfractionalsamplepositions(whiteblockswithlower-caseletters).
Fig.
2.
Theproposedinterpolationforsub3.
THEPROPOSEDALGORITInordertoovercomethedrawbackofthepreviousadaptiveAIFmethodsforKTA,weproposeablockAIFusinglearning-basedSRasaninterpolationtoolforsubpelmotionestimation(seeFig.
2).
Atthefirststep,ofhalf-pelaccuracyareinterpolatedusingthe1ststepdictinary,andtheremainingsub-pels,i.
e.
,n,andoof1/4-pelaccuracyareinterpolatedusingthe2ndstepdictionary.
Fromthisdictionarycanachievequarter-pelmotionestimationandcompensationguaranteeinghighcodingefficiency.
Likeconventionallearning-basedSRalgorithms[optimaldictionariescanbederivedfromsophase,whichisdescribedinthefollowingsubsection.
3.
1Off-lineLearningPhaseFig.
3describestheprocesstoproducetrainingimagestothe1stand2ndstepdictionariesinoffFig.
3,thetrainingimagesofhalfsizeresolutionareproducedfromtheoriginalHRimages.
Here,awell-known5ⅹ5Gaussiananti-aliasingfilter.
The1ststepdictionaryforhalfcompensationisderivedfromhalfages.
The2ndstepdictionaryisgeneratedfromtheHRiagesandthecorrespondingLRimageswhichwererstructedfromthequarter-sizeimagessynthesizedbytheprposedSRbasedonthe1ststepdictionary.
Fig.
3.
Thetrainingimagestogeneratethe1ries.
InterpolationProcess1stStepInputBlockx2interpolationfor1/2-pelME1stStepDictionaryTrainingHRimagesHalfLRimages1stDictionary2DictionaryX……Theproposedinterpolationforsub-pelmotionestimation.
THEPROPOSEDALGORITHMInordertoovercomethedrawbackofthepreviouspicture-adaptiveAIFmethodsforKTA,weproposeablock-adaptivebasedSRasaninterpolationtoolforsub-pelmotionestimation(seeFig.
2).
Atthefirststep,b,h,andjpelaccuracyareinterpolatedusingthe1ststepdictio-pels,i.
e.
,a,c,d,e,f,g,i,k,l,m,pelaccuracyareinterpolatedusingthe2ndstepdictionary.
Fromthisdictionary-driveninterpolation,wepelmotionestimationandcompensationranteeinghighcodingefficiency.
basedSRalgorithms[7-9],theoptimaldictionariescanbederivedfromso-calledlearningphase,whichisdescribedinthefollowingsubsection.
cesstoproducetrainingimagestothestepdictionariesinoff-linelearningphase.
AsinFig.
3,thetrainingimagesofhalf-sizeresolutionandquarter-sizeresolutionareproducedfromtheoriginalHRimages.
5Gaussiankernelisemployedasanstepdictionaryforhalf-pelmotioncompensationisderivedfromhalf-sizeandquarter-sizeim-stepdictionaryisgeneratedfromtheHRim-agesandthecorrespondingLRimageswhichwererecon-sizeimagessynthesizedbythepro-stepdictionary.
Thetrainingimagestogeneratethe1stand2ndstepdictiona-InterpolationProcessInterpolatedBlock(x4)2ndStepx2interpolationfor1/4-pelME2ndStepDictionaryHalf-sizeLRimagesstStepDictionary2ndStepDictionaryx2SRXQuarter-sizereconstructedimages…Quarter-sizeLRimages…2157Fig.
4.
Theoverallprocessofproducingthe1ststepdictionary.
Fig.
4showstheoverallprocessofproducingthe1ststepdictionaryinmoredetail.
First,allpossibleLRandHRpatchpairsofMⅹMsizeareextractedfromseveralquarter-size(LR)andhalf-size(HR)images.
LetiLPandiHPdenotethei-thLRandHRpatchesatthesamespatialposition.
Fig.
5describesanexampleofLRandHRpatcheswhentheirsizesaresetto5ⅹ5.
Inthisfigure,allrectanglesindicateHRpix-elsandgreyrectanglesindicateLRpixels.
EachLRpatchisextractedviaproperoverlappingwithadjacentLRpatches.
Inthecurrentstudy,theM/2pixelsareoverlappedbetweenneighborpatchesinbothdirections.
AninputLRpatchshouldbecomparedwithcandidateLRpatchesinthedictionary,anditsHRpatchissynthesizedusingthehighfrequencyinformationcorrespondingtothecandidateLRpatch(es)withminimumdistance.
Inordertoimprovetheaccuracyofsuchmatchinginthesynthesisphase,LaplacianofLRpatchisemployed[9].
TheLaplacianofeachLRpatchisproducedbyapplyinga3ⅹ3LaplacianoperatortoeverypixelintheLRpatch.
Subsequently,Lapla-cianpatchesarenormalizedforfurtherreliablematching.
LetiLQdenotethenormalizedLaplacianofiLP.
Conventionallearning-basedSRrequiresasmanypatchpairsaspossibletomaintainreliableperformance,whichcausesatremendousmemorycostaswellasasignificantmatchingcomputation.
Therefore,weclustersimilarLRandHRpatchpairs.
WeapplyK-meansclusteringbasedonLQtoallpatchpairs.
Fig.
5.
AnexampleofLRandHRpatchpair.
Fig.
6.
Theclusteringresults.
Asaresult,KLQclustercentersareobtained,andeachclusterisindexedbyitsclustercenter.
NotethatKissignifi-cantlysmallerthanthenumberofentirepatchpairsextractedfromLRandHRtrainingimages.
Fig.
6showstheclusteringresults.
LetjkLP,andjkHP,bethej-thLRandHRpatchesinthek-thcluster.
Then,jkHP,canbecomputedfromjkLP,bythefollowingequation:,),(),(),(1010,,,∑∑===MuMvjkLjkstjkHvuPvuwtsP(1)where(u,v)and(s,t)denotethepixelpositionsintheLRandHRpatches,respectively.
Now,wederiveacommonweightsetkW)1(,i.
e.
,}1,,,0|),({)1(≤≤MvutsvuwkstsuchthatthesquaredsumofinterpolationerrorbyEq.
(1)ismi-nimizedforallLRandHRpatchesinthek-thcluster.
Inor-dertoseeksuchanoptimalweightsetforeachcluster,weemploypopularLMSalgorithm[10].
Thesuperscript(1)ofkW)1(indicatesthe1ststep.
Finally,wecanobtaintheoptim-al1ststepdictionary}1|),{()1()1(KkWQkkL≤≤.
The2ndstepdictionaryisconstructedinthesamewayasthe1ststepdictionary.
Theonlydifferenceisthatthe2ndstepdictionaryistrainedfromtheoriginalHRimagesandthehalf-sizereconstructedimageswhichareup-scaledfromtheircorrespondingquarter-sizeLRimagesusingthe1ststepdic-tionaryasinthefollowingsubsection.
3.
2On-the-flyInterpolationPhaseForsub-pelmotionestimationofeachinputblock,two-stepinterpolationshouldbeperformedonanMⅹMblockbasisbyusingthe1stand2ndstepdictionariesasinFig.
2.
The1ststepinterpolationisdescribedindetailasfollows:Fig.
7describesthe1ststepinterpolationprocessforhalf-pelmotionestimationofanarbitrary4ⅹ4block.
Inthisfig-ure,thered-linerectanglesindicatetheinteger-pixelsofthemotion-compensated4ⅹ4block.
Priortosub-pelmotionestimationofthecurrent4ⅹ4block,thehalf-pelsintheblueregionshouldbeinterpolated.
Inordertointerpolatesuchhalf-pels,nine5ⅹ5LRpatchesaresuper-resolvedinzigzagscanwithoverlappingof3LRpixelsinbothdirections.
Notethatthehalf-pelspixelsonlyintheblueregionneedtobesynthesized.
Half-pelmotionestimationfortheothersizeblockscanbeoperatedsimilarly.
HRTrainingImageLRTrainingImageExtractionofHR-LRPatchPairsNormalizationofLRPatchesK-MeansClusteringCreationofWeight1stStepDictionaryHRpatchLRpatchLRpatchesHRpatches…LRpatchesHRpatches…LRpatchesHRpatchescluster0clusterkclusterKQL0QLkQLK2158Fig.
7.
Theproposedhalf-pelinterpolationprocess.
Here,andindicateinteger-andhalf-pels,respectively.
Thesynthesisprocessofthe1stLRpatchinFig.
7isde-pictedasfollows.
ThenormalizedLaplacianinLQforthein-putLR5ⅹ5blockinLPisfirstderived.
Then,thenearestcandidateLRpatchtoinLQissearchedinthe1ststepdictio-nary.
Inthecurrentstudy,thesumofsquarederrors(SSE)isemployedasthedistortionmeasureformatchingofLapla-cianLRpatches.
LetkbestW)1(betheweightsetcorrespondingtothebest-matchedLaplacianLRpatch.
FromtheinputLRpatch,wecanproducetheinterestinghalf-pelsofthedottedboxinsidethe5ⅹ5HRpatch(seeFig.
7)byusingkbestW)1(andEq.
(1).
Similarly,theremaininghalf-pelscanbeinterpolated.
Forthehalf-pelpositionsintheoverlappingregion,multipleHRpixelvaluessynthesizedbyEq.
(1)areaveraged.
Atthesamefashion,thequarter-pelscanbederivedfromtheinteg-er-andinterpolatedhalf-pelsbyusingthe2ndstepdictionary.
Notethattheproposedalgorithmdoesnothavetotransmitanysideinformationrelatedtofiltercoefficientstothede-coderbecausetheexactfiltercoefficientsofeveryblockcanbeobtainedfromthedictionariesinthedecoder.
4.
EXPERIMENTALRESULTSInordertoevaluatetheproposedalgorithm,ten1920ⅹ1080videosequencesofTable1areused.
Also,six3840ⅹ2160trainingvideosequences,whicharenotincludedinthetestset,areemployedtoderivethe1stand2ndstepdictionaries.
TheproposedinterpolationalgorithmwasimplementedonH.
264KTAsoftwarecalledJM14.
0KTA2.
6.
Forthisexpe-riment,RDoptimizationmodewasoff,CABACwasadoptedforentropycoding,andtheGOPstructurewassettoIPPPP.
Thefirst5framesofeachtestvideosequencewereencodedfor4quantizationparameters(QP),i.
e.
,22,27,32,and37inhighprofile.
Table1.
ThecomparisonintermsofBD_rate(%).
NSAIFBAIFParkjoy-5.
32%-9.
22%Parkscene-6.
01%-9.
53%Crowdrun-8.
11%-13.
41%Bluesky-3.
01%-4.
61%Rollingtomatoes-2.
03%-3.
33%Basketdrive-6.
58%-9.
32%Rushhour-6.
31%-10.
34%Traffic-7.
41%-10.
91%BQTerrace-7.
56%-11.
30%Station-5.
43%-8.
63%Thesearchrangeofinteger-pelmotionestimationwassetto±32.
ThesizeofLRpatchwas5ⅹ5,andthenumberofclus-tersKwas512forboth1stand2ndstepdictionaries.
Table1comparestheproposedBAIFwiththeconvention-alsingle-passNSAIFofKTAand6-tapfilterofH.
264/AVC.
TheywerecomparedintermsofaveragedBD_rate.
Thefixed6-tapfilterofH.
264/AVCwasselectedasabaselinetocomputetheBD-rates.
Forexample,theBAIFprovideshigherBD-rateof5.
3%atmaximumthantheconventionalNSAIFforCrowdrunse-quence.
Ingeneral,theproposedalgorithmshowsmuchbet-tercodingefficiencyforvideosequenceswithcomplextex-turesoredgessuchasCrowdrunthanhomogeneousvideosequencessuchasRollingtomatoes.
Thisisbecausethelearning-basedSRisnormallyveryusefultoaccuratelysyn-thesizetexturesoredges.
Inaddition,Fig.
8comparestheproposedalgorithmwithfixed6-tapfilterofH.
264andAIFofKTAintermsofRD(Rate-Distortion)curves.
Fig.
8.
RDcurvesofseveralalgorithmsforCrowdrunsequence.
The1stLRpatchApartofthe1stHRpatchtobesuper-resolvedThe9thLRpatchAportionofthe9thHRpatchtobesuper-resolved2159(a)(b)(c)(d)Fig.
9.
Apartofthe5thframeofCrowdrun.
(a)Original(b)H.
264(PSNR:35.
92dB,QP:40)(c)NSAIF(PSNR:36.
13dB,QP:37)(d)BAIF(PSNR:36.
41dB,QP:35).
WecanobservethattheBAIFshowsbettercodingper-formanceinhigherbit-rates.
Also,Fig.
9comparesthepro-posedalgorithmwithpreviousworksintermsofsubjectivevisualquality.
TheCrowdrunsequencewasencodedwithproperQPvaluessothatallthealgorithmshavealmostsamebit-rates,andthenapartofthe5thdecodedframewaschosenforcomparison.
WecanseethattheBAIFshowsmuchbettervisualqualitythantheexistingalgorithms.
5.
CONCLUSIONThispaperpresentedablock-adaptiveinterpolationfilteringwhichshowsbetterRDperformanceaswellashighersub-jectivevisualqualitythantheconventionalAIFforsub-pelmotionestimation.
Theproposedalgorithmemployedthelearning-basedSRtomaximizetheinterpolationaccuracy.
Also,theproposedalgorithmdoesnothavetotransmitanysideinformationrelatedtofiltercoefficientsbecausetheex-actfiltercoefficientsofeveryblockcanbederivedfromtheequivalentdictionariesinthedecoder.
SimulationresultsshowthattheproposedalgorithmprovideshigherBD_rateof13.
4%atmaximumthantheconventionalFIRfilterofH.
264/AVC.
ACKNOWLEDGMENTThisresearchwasfinanciallysupportedbytheMinistryofKnowledgeEconomy(MKE)andtheKoreaInstituteforAdvancementofTechnology(KIAT)throughtheHumanResourceTrainingProjectforStrategicTechnology,andwassupportedbytheNationalResearchFoundationofKorea(NRF)grantfundedbytheKoreagovernment(MEST)(No.
2010-0015861).
REFERENCES[1]ITU-TRecommendationH.
264andISO/IEC14496-10(MPEG-4AVC),AdvancedVideoCodingforGenericAudiovisualService,May2003.
[2]Y.
Vatis,B.
Edler,D.
T.
Nguyen,andJ.
Ostermann,"Two-dimensionalnonseparableadaptiveWienerin-terpolationfilterforH.
264/AVC,"ITU-TSG16/Q.
6Doc.
VCEG-Z17,Busan,Korea,April2005.
[3]S.
WittmannandT.
Wedi,"Separableadaptiveinterpo-lationfilter,"ITU-TSG16/Q.
6Doc.
C219,Geneva,Switzerland,June2007.
[4]D.
Rusanovskyy,K.
Ugur,andJ.
Lainema,"Adaptiveinterpolationwithdirectionalfilters,"ITU-TSG16/Q.
6Doc.
VCEG-AG21,Shenzhen,China,Oct.
2007.
[5]ITU-TOutputDocument,"Reportofsubjectivetestresultsofresponsestothejointcallforproposals(CfP)onvideocodingtechnologyforhighefficiencyvideocoding(HEVC),"ITU-TSG16Doc.
JCTVC-A204,Dresden,Denmark,April,2010.
[6]M.
Zhao,M.
Bosma,andG.
deHaan,"Makingthebestoflegacyvideoonmoderndisplays,"J.
Soc.
Inf.
Dis-play,vol.
15,no.
1,pp.
49-60,2007.
[7]W.
Freeman,T.
Jones,andE.
Pasztor,"Example-basedsuper-resolution,"IEEEComputerGraphicsandAp-plications,vol.
40,no.
1,pp.
23-47,2000.
[8]W.
FanandD.
Yeung,"Imagehallucinationusingneighborembeddingovervisualprimitivemanifolds,"IEEEProc.
CVPR,2007.
[9]S.
C.
JeongandB.
C.
Song,"Noise-robustsuper-resolutionbasedonaclassifieddictionary,"JournalofElectronicImaging,Dec.
2010.
[10]S.
Haykin,AdaptiveFilterTheory,Chap.
9,3rded.
,PrenticeHall,1996.
2160