Fast H.264 AVC FRExt Intra Coding Using Belief Pro

Copyright(c)2010IEEE.Personaluseispermitted.Foranyotherpurposes,[e-mail]edforpublicationinafutureissueofthisjournal,buthasnotbeenfullyedited.Contentmaychangepriortofinalpublication.1FastH.264/AVCFRExtIntraCodingUsingBeliefPropagationSimoneMilani,Member,IEEEAbstractIntheH.264/AVCFRExtcoder,thecodingperformanceofIntracodingsignificantlyovercomesthepreviousstillimagecodingstandards,likeJPEG2000,thankstoamassiveuseofspatialprediction.Unfortunately,theadoptionofanextensivesetofpredictorsinducesasignificantincreaseofthecompu-tationalcomplexityrequiredbytherate-distortionoptimizationroutine.ThepaperpresentsacomplexityreductionstrategythataimsatreducingthecomputationalloadoftheIntracodingwithasmalllossinthecompressionperformance.Theproposedalgorithmreliesonselectingareducedsetofpredictionmodesaccordingtotheirprobabilities,whichareestimatedadoptingaBelief-Propagationprocedure.Experimentalresultsshowthattheproposedmethodpermitssavingupto60%ofthecodingtimerequiredbyanexhaustiverate-distortionoptimizationmethodwithanegligiblelossinperformance.Moreover,itpermitsanaccuratecontrolofthecomputationalcomplexityunlikeothermethodswherethecomputationalcomplexitydependsonthecodedsequence.IndexTermsfastIntracoding,videocoding,H.264/AVC,beliefpropagation,rate-distortionoptimization.EDICSCategories:COM-LOC,COM-MMCI.INTRODUCTIONIntheH.264/AVCstandardizationprocessthecompressionperformanceofIntracodingwassignifi-cantlyimprovedbytheadoptionofspatialpredictioninIntraframes.Thepixelsofthecurrentblockarepredictedusingthereconstructedpixelsofneighboringblocksinterpolatedalongdifferentorientations,whichresultcloselyrelatedtothecharacteristicsoftheimagecorrelation[1].InthefirstversionoftheSimoneMilaniiswiththeDept.ofInformationEngineering,UniversityofPadova,viaGradenigo6/B,35131Padova,Italy.Phone:[phone]Fax:[phone].E-mail:[e-mail]17,2010DRAFTCopyright(c)2010IEEE.Personaluseispermitted.Foranyotherpurposes,[e-mail]edforpublicationinafutureissueofthisjournal,buthasnotbeenfullyedited.Contentmaychangepriortofinalpublication.2H.264/AVCstandard,thespatialpredictionislimitedtoeitherblocksof4×4pixelsorwholemacroblocks(MBs)of16×16pixels.IntheFRExtextensionofthestandard,blocksof8×8pixelsareconsideredtoo.Thisstrategy,togetherwithanimprovedrate-distortionoptimizationandaneffectivearithmeticcodingengine,permitsobtainingahighercompressiongainwithrespecttothepreviousimagecodingstandardslikeJPEG2000(see[2]),butunfortunately,theadditionaloperations,whicharerequiredtofindthemostappropriatespatialpredictionandblocksize,significantlyincreasethefinalcomputationalcomplexity.ThisdrawbackmakestheadoptionofH.264IntracodingmoretroublesomefortraditionalIntra-onlyvideocodingsuchasvideosurveillancesystems,wherereal-timevideoprocessingisrequiredandthecomputationalloadmustbeconstrained.Inordertoovercomethisproblem,awidevarietyofcomplexityreductionstrategies,togetherwiththeintroductionofnovelhardwareaccelerators,havebeenproposedinliterature.In[3]Panetal.proposeafastIntrapredictionalgorithmthatextractstheimagefeaturesusingSobeledgeoperatorsandchoosesthepredictoraccordingtotheirstatistics.Inasimilarway,theapproachesin[4],[5]extractthedirectionalfeaturesofeachframeandusethemtoestimatethemostprobablepredictionmodes.Thesolutionsproposedin[6],[7]evaluatethedistortionproducedbypredictioninthetransformdomain,whilein[8]Kimetal.jointlyextractthefeaturesofeachblockfrombothpixelsandtransformcoefficients.Inaddition,temporalcorrelationexistingbetweenadjacentframescanbeusedtoo,asitisshownin[9].Manyapproachesemployearly-terminationdecisioninordertoreducetheamountofcomputation[10].Thismakesthecomputationalcomplexityvarysignificantlyaccordingtotheprocessedvideosequence(see[11]asanexamplewheretherelativereductionofcodingtimevariesfrom40%to70%),andtherefore,an“apriori”estimationoftheresultingcostisnotpossible.Atthesametime,theperformanceofthealgorithmvariesaccordingtothecodedsequencelikeinthecaseof[12]whereamachinelearningalgorithmisusedtoselectthebestpredictionmodeamongareducedsetofcandidates.Withrespecttothesemethods,thedesignofacomplexityreductionstrategythatpermitscontrollingtheamountofrequiredcomputationprovidesseveraladvantages,suchas•thepossibilityofadaptingthealgorithmtodeviceswithdifferentcomputationalcapabilitiesandpowersupply;•anaccurateestimationoftheautonomyofbattery-poweredcodingdevices;•thepossibilityofenablingpowersavingconfigurationsthatgraduallyreducethecomputationalcomplexity(atthecostofaworserate-distortionoptimization)accordingtotheremainingbatterycharge.June17,2010DRAFTCopyright(c)2010IEEE.Personaluseispermitted.Foranyotherpurposes,[e-mail]edforpublicationinafutureissueofthisjournal,buthasnotbeenfullyedited.Contentmaychangepriortofinalpublication.3Fig.1.BlockdiagramoftheH.264/AVCcoder.Theso