numav

FRAMELEVELRATECONTROLFORH.
264/COLOR:#000000;BACKGROUND-COLOR:#ffff00">AVCWITHNOVELRATE-QUANTIZATIONMODELSudengHu1,HanliWang2,,SamKwong1,,TiesongZhao11DepartmentofComputerScience,CityUniversityofHongKong,HongKong,China2DepartmentofComputerScience,TongjiUniversity,Shanghai201804,ChinaEmail:sudenghu@gmail.
com,hanli.
wang@ieee.
org,cssamk@cityu.
edu.
hk,ztiesong2@student.
cityu.
edu.
hkABSTRACTInthispaper,aframelevelratecontrolalgorithmisproposedwithanovelRate-Quantization(R-Q)modelforH.
264/COLOR:#000000;BACKGROUND-COLOR:#ffff00">AVC.
Firstly,atwo-stageratecontrolschemeisadoptedtodecou-pletheinter-dependencybetweenRateDistortionOptimization(RDO)andratecontrol.
Secondly,inordertopredicttheframecomplexityaccurately,insteadoftheMeanAbsoluteDifference(MAD)oftheresidualsignal,bitsinformationintheRDO-basedmodedecisionprocessisemployedtopredicttheframecomplexity.
Thirdly,aself-adaptiveexponentialR-Qmodelisproposedforratecontrol.
ExperimentalresultsrevealthattheproposedR-Qmodelcanestimatetheactualoutputbitsverywell,andthenovelratecontrolschemehasexcellentperfor-mancebothinbitrateaccuracyandcodingefciencyascom-paredtoJVT-W043andtheFixedQptoolintheJointScalableVideoModelreferencesoftware.
Keywords—Ratecontrol,H.
264/COLOR:#000000;BACKGROUND-COLOR:#ffff00">AVC,R-Qmodel.
1.
INTRODUCTIONRateControl(RC)isemployedtoregulateoutputbitstreamaccordingtothebandwidthlimitation,meanwhileaimingtooptimizethevisualquality.
Usually,thispurposeisachievedthroughtwosteps.
Attherststep,efcientbitbudgetisdis-tributedforeachcodingunitsuchasaframeoramacroblock(MB).
Atthesecondstep,aproperquantizationparameterischosentoachievethisbitbudget.
Inordertoreachthetargetbitaccuratelyatthesecondstep,severalRCalgorithmshCOLOR:#000000;BACKGROUND-COLOR:#ffff00">avebeendevelopedtomodeltheR-Qrelationship.
In[1],basedontheassumptionthatDCTcoef-cientsareLaplaciandistributed,asecond-orderquadraticmodelisderived.
Howeveritdoesnottakeframecomplexityintocon-sideration.
In[2],theclassicquadraticmodelisappliedandtheframecomplexityisconsideredbypredictingtheMADofen-codingframe,whichisimplementedasthenon-normativeRCalgorithminH.
264/COLOR:#000000;BACKGROUND-COLOR:#ffff00">AVCJointModel(JM)referencesoftwares.
ThisworkissupportedbytheHongKongResearchGrantsCouncilGeneralResearchFund,underProjects9041236(CityU114707)and9041353(CityU115408),andtheGermanAlexandervonHumboldtResearchFellowship.
*Cor-respondingauthor.
In[3,4,5],simpliedlinearR-QmodelsareproposedwhichhCOLOR:#000000;BACKGROUND-COLOR:#ffff00">avecomparablecodingefciencytothequadraticmodel.
In[6],basedontheassumptionthatDCTcoefcientsareCauchydistributed,anexponentialmodelisdeveloped.
In[7],therela-tionshipbetweenrateandquantizationisbuiltthroughρ,whichisthepercentageofzerosamongthequantizedtransformcoef-cients.
Asstatedabove,themodelparametersincludingtheframecomplexityintermsofMADandzerocoefcients'percent-ageρarecrucialtoconstructaccurateR-Qmodels.
However,itisnottrivialtoobtaintheseparameters.
RegardingMAD,thetraditionalwaytocharacterizeitisnotveryeffective,sinceMADisindirectlyrelatedtothenumberofoutputbitsandthusitgenerallydoesnotholdstrongrelationshipwiththeactualoutputbitrate.
Moreover,duetotheinter-dependencybetweenRCandRDO,theMADofacodingframeisusuallypredictedfromitspreviousframe,whichfurtherdecreasestheaccuracyofthecomplexityestimation.
Ontheotherhand,asfarastheρ-domainmethodisconcerned,althoughithasarelativelycloserelationwiththeconsumedbitrate,moreDiscreteCosineTransform(DCT)andquantizationcomputationsarerequiredformodelparametergeneration.
Inthispaper,inordertoimprovetheR-Qmodelperfor-mance,insteadofutilizingMADorρ,theinformationofbitconsumptionintheRDOprocessisemployedtopredictthecomplexityofacodingframe.
Withthisnovelcomplexitymea-sure,animprovedexponentialR-QmodelisdevelopedforRCrealization.
Therestofthispaperisorganizedasfollows.
Theframecomplexitypredictionwithtwo-stageRCschemeispre-sentedinSection2.
Then,theself-adaptiveexponentialR-QmodelisintroducedinSection3,andtheproposedoverallRCalgorithmisdescribedtherein.
InSection4,experimentalre-sultsarepresentedtoillustratetheefciencyoftheproposedRCalgorithm.
Finally,conclusionsaregiveninSection5.
2.
MEASUREMENTOFFRAMECOMPLEXITY2.
1.
Two-StageRCSchemeTheRDO-basedmodedecisionimplementedinH.
264/COLOR:#000000;BACKGROUND-COLOR:#ffff00">AVCim-provesthecodingefciencysignicantlyascomparedwithpre-978-1-4244-7493-6/10/$26.
00c2010IEEEICME2010viouscodingstandards.
Inmodedecision,thebestcodingmodeisselectedwiththeminimalRate-Distortion(R-D)costaccord-ingto[8]:J=Dmode+λRmode,(1)whereDmodeandRmodearethedistortionandthenumberofoutputbitsofaMBencodedinaspecicINTERorINTRAmode;λistheLagrangianmultiplierdependingonquantiza-tionparameterQp.
ForeachMB,allpossibleencodingmodesaretriedandtheencodingmodewiththeminimalRDcostbyminimizingJinEq.
(1)ischosenasthebest.
TheRDO-basedmodedecisionprocesscausestheso-called"Chicken-and-Egg"dilemma[9]inratecontrolforH.
264/COLOR:#000000;BACKGROUND-COLOR:#ffff00">AVC,becauseQpisrequiredbeforeRDOprocessbutresidualsig-nalanditsrelatedinformationisunCOLOR:#000000;BACKGROUND-COLOR:#ffff00">availablebeforethatstage.
In[10],atwo-Qpschemeisproposedtodecoupletheinter-dependencyproblembetweenRDOandratecontrol:oneisusedformodedecisionandtheotherisemployedinquanti-zation.
Theoretically,toreachtheR-Doptimization,itrequiresthesetwoQpvaluestobeequal.
However,itisobservedthatasmallmismatchbetweenthesetwoQpvalueswillnotdecreasethecodingefciencysignicantly.
Inspiredby[10],atwo-stageframelevelratecontrolschemeisdevelopedinthiswork.
Attherststage,apredeneQpdenotedasQp1isadoptedinRDOprocessforallMBsofaframe.
Atthesecondstage,theotherQpdenotedasQp2iscalculatedforquantizationtogeneratethenumberoftargetbits.
Beforeencodingtheithframeinasequence,theQp1(i)issetaccordingtothepreviousQp2value.
Morespecically,itissettoQp(i)whichisupdatedasQp(i)=wq·Qp2(i1)+(1wq)·Qp(i1),(2)wherewqistheweightingparameter,whichissetto0.
7basedonexperiments.
ForQp2atthesecondstage,itiscalculatedaccordingtotheR-QmodelthatwillbediscussedinSection3.
2.
2.
ProposedFrameComplexityMeasurementDuetothevariantcontentofdifferentframes,theframesevenencodedwiththesameQpoftenproducequitedifferentnum-bersofoutputbits.
InordertobuildanaccurateR-Qmodel,theMADofresidualsignalisusuallyemployedtodescribetheframecomplexityinnon-normativeRCalgorithmsofvideocodingstandards.
ThefollowingmodelisconsideredastheclassicR-Qmodel:R=α1·MADQstep+α2·MADQ2step,(3)whereα1andα2aremodelparametersandQstepisthequan-tizationstepsize.
Inthismodel,MADisutilizedtomodelthelinearrelationbetweencomplexityandthenumberofoutputbits.
However,asillustratedinFig.
1,thiskindoflinearrela-tionshipisnotverystrongandconsequentlyitwilldecreasetheaccuracyoftheR-Qmodel.
Inordertoestimatetheframecomplexityaccurately,insteadofusingMAD,bitsinformationCOLOR:#000000;BACKGROUND-COLOR:#ffff00">availableintheRDO-based1.
71.
81.
9220040060080010001200MADBits(a)22.
533.
5405000100001500020000MADBits(b)Fig.
1.
TherelationshipbetweenMADandthenumberofout-putbitsforframes.
(a)"Container"sequence,QCIFformat,241frames,Qp=28.
(b)"Foreman"sequence,QCIFformat,241frames,Qp=28.
modedecisionprocessisinvestigatedtopredicttheframecom-plexity.
Thisisbecauseatthemodedecisionstage,theapprox-imatenumberofoutputbitsofeachMBwillbeestimatedandthusCOLOR:#000000;BACKGROUND-COLOR:#ffff00">available.
LetRhbest(i)andRtbest(i)denotetheapproxi-matenumberofheaderbitsandtexturebitsforencodingtheithMBinthemodedecisionprocess,respectively,thecomplexityofthecorrespondingframeisdenedasC=Ch+Ct,(4)wheretheheadercomplexityChandtexturecomplexityCtareCh=Ni=1Rhbest(i),Ct=Ni=1Rtbest(i),(5)inwhichNisthenumberofMBsinsideaframe.
Then,theproposedframecomplexityCisusedtoderivethequantizationparameterQp2forthesubsequentquantizationprocesstogen-eratetheactualoutputbits.
InordertoshowtheefciencyoftheproposedframecomplexityC,therelationshipbetweenCandtheactualnumberofoutputbitsisillustratedinFig.
2,whereitcanbeobservedthatthereisaquasi-linearrelationshipbetweenCandtheactualnumberofoutputbits,andthusCisbetterthanMADforcomplexitypredicationpurposes.
050010001500050010001500CBits(a)00.
511.
522.
5x10400.
511.
522.
5x104CBits(b)Fig.
2.
TherelationshipbetweenCandactualoutputbitsforframes.
(a)"Container"sequence,QCIFformat,241frames,Qp1=Qp2=28.
(b)"Foreman"sequence,QCIFformat,241frames,Qp1=Qp2=28.
3.
PROPOSEDRCALGORITHMWITHNOVELR-QMODEL3.
1.
ExponentialR-QModelBeforequantizationandentropycoding,thetwo-dimensionalDCTisappliedtoreducethespatialredundancy.
Theproba-bilitydistributionoftheDCTcoefcientsbecomeimportanttobuildareasonableR-Qmodel.
SeveraldistributionmodelsareproposedtomodeltheactualdistributionofDCTcoefcients,andamongthemtheCauchydistributionisreportedtohCOLOR:#000000;BACKGROUND-COLOR:#ffff00">avebet-teraccuracythanothermodelsin[6].
TheCauchydistribution-basedR-QmodelisexpressedbyRt=a·Qβstep,(6)whereaisthecomplexityrelatedparameterandβisthemodelparameterassociatedwithDCTcoefcientsdistributionchar-acteristics.
Inthismodel,acouldbeupdatedintheencodingprocess,whileβislimitedtoasetofconstantvaluesaccordingtodifferentframetypes,e.
g.
,{0.
75,0.
8,0.
85}forIframe,{1.
2,1.
4,1.
6}forPframe,{1.
6,1.
8,2.
0}forBframe.
Intheproposedtwo-stageRCalgorithm,aftertherststage,themodelparameteracanbeestimatedaccordingtotexturecomplexityasa=Ct·Qβstep1,(7)whereQstep1isthequantizationstepsizecorrespondingtoQp1usedattherststage.
Therefore,themodeloftexturebitswith0204060801001200500100015002000250030003500Qstep2HeaderBitsQp1=15Qp1=25Qp1=35Fig.
3.
TherelationshipbetweenthenumberofheaderbitsandQstep2.
ThecorrespondingQp2ischangedfrom15to45whileQp1issetto15,25,and35,respectively.
respecttoQstep2(whichisthequantizationstepsizecorre-spondingtoQp2usedatthesecondstage)canbewrittenasRt(Qstep2)=CtQβstep1Qβstep2.
(8)SincethenumberofheaderbitsmainlycomesfromMotionVectors(MVs),modetypeandetc.
,whicharedecidedattherststagebyQp1.
Therefore,itisalmostnotaffectedbyQstep2,asshowninFig.
3.
Consequently,thetotalnumberofframebitscanbemodeledasR(Qstep2)=CtQβstep1Qβstep2+Ch,(9)whereRisthenumberoftargetbitsforthecurrentframe.
TheclassicbitallocationschemeinJVT-W043[11]isadoptedinthecurrentworktocalculatethenumberoftargetbitsR.
Inthismodel,MADisreplacedbyCtasframecomplexityforR-Qmodel.
Figure4showstherelationbetweentheactualnumberofoutputbitsandthepredictednumberofbitsbytheproposedmodelwithdifferentβvalues.
InFig.
4,itisobviousthatβiscrucialtotheaccuracyofthemodel.
However,whenaproperβischosen,thepredictednumberofbitsmatchestheactualnumberofoutputbitswellinarangeofQstepvalues.
3.
2.
ModelParameterUpdateIntheR-QmodelinEq.
(9),theparameterβisrelatedtodistri-butionofACDCTcoefcients.
Usually,itisapredenedcon-stantparameter.
However,thedistributionofactualACDCTcoefcientsofdifferentframesvariessignicantlyindifferentsequencesorevenindifferentframesofthesamesequence.
Moreover,asshowninFig.
4,theβvalueiscriticaltothemodel.
Inthesequel,itisdesirabletoupdateβaccordingtotheframecharacteristicsintheencodingprocess.
Aftercod-ingtheithframe,theactualβforthisframecanbecalculatedaccordingtoEq.
(8)asβ(i)=lnRt(i)/Ct(i)ln(Qstep1(i)/Qstep2(i)),(10)0204060801001201400.
511.
522.
533.
544.
55x104Qstep2BitsActualbeta=4.
2beta=3.
0beta=1.
2Fig.
4.
TheactualandpredictedR-Qcurves.
Qp2ischangedfrom25to46,andβissetto1.
2,3.
0,and4.
2fortheproposedR-Qmodel.
whereRt(i)isthenumberofactualtextualoutputbitsfortheithframe.
Sinceneighboringframesgenerallyexhibitsimilarcharac-teristicsinavideosequenceduetothetemporalcorrelation,weassumethesimilarityinβfornearbyframes.
Asaresult,aftercodingtheithframe,βisupdatedandpredictedforthe(i+1)thframeasβ(i+1)=wβ·β(i)+(1wβ)·β(i),(11)whereβ(i)isthepredictedvaluefortheithframe;wβistheweightingparameterwiththetypicalvalueof0.
7inthiswork.
3.
3.
OverallRCAlgorithmBasedontheaboveanalyses,theproposedoverallRCalgorithmincludestwostagesattheframelevel,withthefollowingstep-by-stepdescriptions.
3.
3.
1.
StageOneStep1.
Qp1iscalculatedasinEq.
(2).
Step2.
ThecalculatedQp1isappliedtoallMBsinthemodedecisionprocessforthecurrentframe.
Step3.
CtandCharerecorded.
Step4.
ThemodelparameterβispredictedaccordingtoEq.
(11).
3.
3.
2.
StageTwoStep5.
Qp2iscalculatedintheproposedR-Qmodelbasedonthenumberoftargetbits.
Step6.
Qp2isclippedasQp2=max{Qp13,min{Qp1+3,Qp2}}.
(12)Step7.
Qp2isappliedinthequantizationprocessforallMBsinthecurrentframe.
ThenumberofactualtextureoutputbitsRtisrecorded,andtheactualvalueofβisupdatedaccordingtoEq.
(10).
Step8.
Finishencodingthecurrentframe,andprocessthenextframe.
4.
EXPERIMENTALRESULTSTheproposedRCalgorithmisimplementedintheH.
264/SVC-basedJointScalableVideoModel(JSVM)referencesoftware9.
17[12]whereonlyasinglelayerisencodedtocomplywithH.
264/COLOR:#000000;BACKGROUND-COLOR:#ffff00">AVCencodingconditions.
Thetestframerateissetto30fps.
TheGOPsizeissetto4,wherethreeB-framesareinsertedbetweenI/Pframes.
AllvideosequenceshCOLOR:#000000;BACKGROUND-COLOR:#ffff00">ave241picturestobecoded.
0100200300400500600343638404244464850BitRate(kb/s)PSNR(dB)W043FixedQpProposed(a)100200300400500600700800900303234363840BitRate(kb/s)PSNR(dB)W043FixedQpProposed(b)Fig.
5.
R-Dcurves.
(a)"Silent"sequenceinQCIFformat.
(b)"Table"sequenceinCIFformat.
ThealgorithmJVT-W043[11]andtheFixedQptoolinJSVMareutilizedforcomparisonwiththeproposedalgorithm.
InJVT-W043,theclassicquadraticalmodelisappliedandtheFixedQptoolisamultiple-passRCalgorithmwherealogarith-micsearchisappliedtondaproperQptomeetthetargetbitrate.
VariousbenchmarkvideosequencesaretestedinbothQCIFandCIF.
Table1.
ComparisonofbitrateaccuracyandPSNR.
JVT-W043FixedQpProposedSequenceTRPSNRERPSNREBPBBRPSNREBPBB(kb/s)(kb/s)(dB)(%)(kb/s)(dB)(%)(dB)(kb/s)(kb/s)(dB)(%)(dB)(kb/s)Akiyo6467.
1344.
564.
8960.
3845.
405.
660.
97-15.
463.
7444.
940.
411.
20-18.
0(QCIF)128135.
7048.
156.
02127.
0049.
010.
78127.
6048.
660.
31256266.
9051.
684.
26251.
6052.
061.
72255.
4053.
020.
23512522.
8056.
702.
11496.
5056.
823.
03500.
4057.
982.
27Foreman6466.
6532.
844.
1463.
0532.
841.
480.
44-7.
463.
7233.
050.
440.
46-7.
8(QCIF)128132.
1036.
703.
20126.
8036.
900.
94127.
3036.
980.
55256262.
6040.
622.
58255.
8040.
970.
08254.
8040.
950.
47512521.
5044.
651.
86509.
3044.
910.
53510.
8044.
800.
23Paris6466.
9332.
924.
5858.
3832.
378.
780.
26-3.
563.
9432.
870.
090.
34-4.
6(QCIF)128133.
5037.
764.
30118.
0037.
127.
81128.
1037.
810.
08256265.
5042.
673.
71257.
3042.
680.
51255.
9042.
770.
04512528.
4047.
633.
20508.
3047.
610.
72511.
8047.
700.
04Silent6467.
0335.
544.
7361.
6835.
343.
630.
32-4.
864.
6535.
771.
020.
78-11.
4(QCIF)128133.
1040.
053.
98126.
1039.
831.
48128.
9040.
750.
70256265.
3044.
413.
63245.
0044.
424.
30246.
3044.
873.
79512523.
8048.
962.
30504.
8048.
981.
41511.
3049.
170.
14Table6467.
1234.
034.
8866.
2234.
233.
47-0.
163.
264.
1734.
010.
270.
29-4.
9(QCIF)128134.
2038.
084.
84125.
1037.
382.
27128.
1038.
030.
08256267.
5041.
914.
49252.
6041.
201.
33244.
4041.
574.
53512565.
5045.
7110.
45508.
8045.
560.
62512.
9046.
010.
18COLOR:#000000;BACKGROUND-COLOR:#ffff00">Average4.
202.
530.
37-5.
60.
790.
62-0.
9Mobile128137.
5024.
977.
42123.
4024.
383.
590.
14-3.
2128.
5024.
930.
390.
20-4.
5(CIF)256268.
4027.
954.
84249.
6027.
782.
50255.
6027.
910.
16512540.
1030.
695.
49491.
4030.
514.
02510.
7030.
680.
25768802.
1032.
444.
44769.
3032.
430.
17765.
5032.
420.
33Table128134.
6031.
355.
16122.
9030.
923.
98-0.
112.
6128.
1031.
170.
080.
31-6.
5(CIF)256268.
5034.
304.
88249.
3033.
822.
62256.
0034.
410.
00512536.
7037.
394.
82495.
1036.
903.
30512.
0037.
550.
00768803.
2039.
364.
58801.
4039.
354.
35768.
9039.
560.
12Silent128134.
7033.
805.
23120.
3033.
476.
020.
28-5.
3129.
5034.
061.
170.
50-9.
3(CIF)256268.
0037.
184.
69256.
0037.
200.
00258.
8037.
431.
09512531.
0040.
633.
71508.
1040.
770.
76513.
5041.
070.
29768795.
8042.
813.
62777.
7042.
941.
26771.
3043.
240.
43News128136.
3036.
936.
48126.
8036.
780.
940.
32-6.
3128.
1036.
830.
080.
55-10.
6(CIF)256274.
0040.
507.
03248.
2040.
423.
05256.
1040.
820.
04512543.
2043.
836.
09511.
0043.
860.
20511.
9044.
140.
02768809.
7045.
695.
43771.
5045.
490.
46765.
9045.
820.
27COLOR:#000000;BACKGROUND-COLOR:#ffff00">Average5.
242.
330.
28-3.
60.
290.
52-8.
7Inordertoevaluatetheaccuracyofbitrateachievement,thefollowingmeasurementisusedasE=|RtRo|Rt*100%,(13)whereRtandRoarethetargetbitrateandactualoutputbitrate,respectively.
ThemismatchofbitrateatdifferenttesttargetbitsbythecomparativealgorithmsarealsopresentedinTables1.
AsshowninTables1,theproposedalgorithmcanachievemuchbetterperformancethanJVT-W043andFixedQpinbitrateaccuracy.
SincetheoutputbitrateofthesethreeRCalgorithmsarenotmatchedexactly,theperformancesofBD-PSNR(denotedbyBP)andBD-BR(denotedbyBB)[13]areemployedinourex-perimentsforafaircomparison.
JVT-W043issetasthebench-mark,andbothoftheFixedQptoolandtheproposedalgorithmarecomparedwithJVT-W043inBPandBB.
Fromtheexper-imentalresultsinTables1,theFixedQptoolisabout0.
37dBinQCIFand0.
28dBinCIFbetterthanJVT-W043onCOLOR:#000000;BACKGROUND-COLOR:#ffff00">average,whiletheproposedalgorithmisabout0.
62dBinQCIFand0.
52dBinCIFbetterthanJVT-W043.
TheRDcurvesfortwobenchmarkvideosequencesaregiveninFig.
5,whichillustratetheRDperformanceoftheproposedRCalgorithmisbetterthanbothJVT-W043andFixedQptoolinawiderangeofbitrates.
5.
CONCLUSIONInthispaper,anovelR-Qmodelisproposedwithatwo-stageRCschemeforH.
264/COLOR:#000000;BACKGROUND-COLOR:#ffff00">AVC.
Thetwo-stageRCschemeisabletodecoupletheinter-dependencybetweenRCandtheRDOpro-cess.
Inaddition,theproposedR-Qmodelutilizesthebitinfor-mationCOLOR:#000000;BACKGROUND-COLOR:#ffff00">availableintheRDOprocessforeffectivelypredictingthefamecomplexity,whichismorerobustandeffective.
TheexperimentalresultsdemonstratethattheRDperformancescanbeimprovedsignicantlywiththeproposedRCalgorithmascomparedtoJVT-W043andtheFixedQptoolunderthetestconditions.
6.
REFERENCES[1]T.
ChiangandY.
-Q.
Zhang,"ANewRateControlSchemeUsingQuadraticRateDistortionModel,"IEEETrans.
Cir-cuitsSyst.
VideoTechnol.
,Vol.
7,No.
1,pp.
246-250,Feb.
1997.
[2]Z.
Li,F.
Pan,K.
P.
Lim,G.
Feng,X.
Lin,andS.
Rahardja,"AdaptiveBasicUnitLayerRateControlforJVT",Doc.
JVT-G012-r1,Pattaya,Thailand,Mar.
2003.
[3]Y.
Liu,Z.
G.
Li,andY.
C.
Soh,"ANovelRateCon-trolSchemeforLowDelayVideoCommunicationofH.
264/COLOR:#000000;BACKGROUND-COLOR:#ffff00">AVCStandard",IEEETrans.
CircuitsSyst.
VideoTechnol.
,Vol.
17,No.
1,pp.
68-78,Jan.
2007.
[4]S.
Ma,W.
Gao,andY.
Lu,"Rate-DistortionAnalysisforH.
264/COLOR:#000000;BACKGROUND-COLOR:#ffff00">AVCVideoCodingandItsApplicationtoRateControl",IEEETrans.
CircuitsSyst.
VideoTechnol.
,Vol.
15,No.
12,pp.
1533-1544,Dec.
2005.
[5]H.
WangandS.
Kwong,"Rate-DistortionOptimizationofRateControlforH.
264withAdaptiveInitialQuantiza-tionParameterDetermination",IEEETrans.
CircuitsSyst.
VideoTechnol.
,Vol.
18,No.
1,pp.
140-145,Jan.
2008.
[6]N.
Kamaci,Y.
Altinbasak,andR.
M.
Mersereau,"FrameBitAllocationfortheH.
264/COLOR:#000000;BACKGROUND-COLOR:#ffff00">AVCVideoCoderviaCauchyDensity-basedRateandDistortionModels",IEEETrans.
CircuitsSyst.
VideoTechnol.
,Vol.
15,No.
8,pp.
994-1006,Aug.
2005.
[7]Z.
HeandS.
K.
Mitra,"OptimumBitAllocationandAccurateRateControlforVideoCodingviaρ-DomainSourceModeling",IEEETrans.
CircuitsSyst.
VideoTech-nol.
,Vol.
12,No.
10,pp.
840-849,Oct.
2001.
[8]G.
J.
SullivanandT.
Wiegand,"Rate-DistortionOptimiza-tionforVideoCompression",IEEESignalProcess.
Mag.
,Vol.
15,No.
6,pp.
23-50,Nov.
1998.
[9]Z.
G.
Li,F.
Pan,K.
P.
Lim,andS.
Rahardja,"AdaptiveRateControlforH.
264",inProc.
IEEEInt.
Conf.
ImageProcess.
,pp.
449-452,Oct.
2004.
[10]D.
Kwon,M.
Shen,andC.
-C.
J.
Kuo,"RateControlforH.
264VideowithEnhancedRateandDistortionmodels",IEEETrans.
CircuitsSyst.
VideoTechnol.
,Vol.
17,No.
5,pp.
517-529,May2007.
[11]A.
LeontarisandA.
M.
Tourapis,"RateControlfortheJointScalableVideoModel(JSVM)",Doc.
JVT-W043,SanJose,USA,Apr.
2007.
[12]"JointScalableVideoModelJSVM9.
17SoftwarePack-age",CVSserverforJSVMreferencesoftwares,Mar.
2009.
[13]G.
Bjontegaard,"CalculationofCOLOR:#000000;BACKGROUND-COLOR:#ffff00">AveragePSNRDiffer-encesbetweenRD-Curves",Doc.
VCEG-M33,Austin,USA,Apr.
2001.