资源描述：
《rate-distortion analysis for vector quantization based on a variable block-size classificat》由会员上传分享，免费在线阅读，更多相关内容在教育资源-天天文库。
1、Rate-distortionanalysisforvectorquantizationbasedonavariableblock-sizeclassi cationmodelM.H.Lee,K.N.NganandG.A.CrebbinDepartmentofElectricalandElectronicEngineeringTheUniversityofWesternAustraliaNedlands,WA6907,AustraliaABSTRACTVectorquantization(VQ)basedona xedblock-sizeclassi cati
2、on(FBSC)model,whichisknownasclassi edVQ(CVQ),o ersausefulsolutionfortheedgedegradationproblemofconventionalimageVQ.Inourpreviouswork,wehavedevelopedaVQtechniquebasedonavariableblock-sizeclassi cation(VBSC)model,inwhichanimageissegmentedintoblocksofvarioussize,andeachsegmentedregioni
3、sencodedatadi erentrateaccordingtoitslevelofdetail.Thelow-detailregionsoftheimageconsistofvariablesizeblocksandareencodedatverylowbitrateswithlittleperceptualdegradation.High-detailregions,whichareisolatedintothesmallestblocks,areclassi edintovariousedgesofwhicheachisseparatelyencod
4、ed.Inthispaper,arate-distortionfunction(RDF),R(D),ispresentedforaVBSCmodel.WeobtainatheoreticalR(D)boundontheperformanceofVQbasedonaVBSCmodel.ItistheoreticallyprovedthattheR(D)boundoftheVBSCmodelislowerthanthoseoftheGaussianmodelandtheFBSCmodel.WealsoexperimentallyevaluateaRDFforthe
5、VBSCmodelandcomparewiththetheoreticalRDF.Thereisagapofabout0.1bppbetweenthetheoreticalRDFandtheexperimentalRDFinVBSCmodel-basedVQcoding.Weexpectthatthisgapcanbereducedbysubsequentlyemployinganentropycoder.Keywords:imagecoding,vectorquantization,rate-distortionfunction1.INTRODUCTION1
6、SinceRamamurthiandGershoproposedavectorquantization(VQ)techniquebasedonaclassi cationmodel,2?5whichisknownasclassi edVQ(CVQ),severalpapershavedemonstratedthatCVQo ersanexcellentsolutionfortheedgedegradationproblemofconventionalimageVQ.Theuseofasmallerblocksizeinblockcodinghasbeenkno
7、wntoleadtobettersubjectivequality,andmoststudiesinVQandCVQhavefocussedondevelopingalgorithmsthatuse xedsmall-sizedblocks,typically44.However,theblocksizeinVQistoosmalltotakeadvantageofthefactthatfewerbitsarerequiredtoencodelow-detailregions,whichusuallyoccupylargeareasinanimage.Rec
8、ently,therehasbeenm
