電子信息工程外文翻譯-使用自適應(yīng)預(yù)測和自適應(yīng)算術(shù)編碼的有損圖像的無損壓縮

附錄B外文參考文獻(xiàn)LosslessImageCompressionwithLossyImageUsingAdaptivePredictionandArithmeticCodingSeishiTakaandMikioTakagiInstituteofIndustrialScience,UniversityofTokyoAbstractLosslessgrayscaleimagecompressionisnecessaryinmanypurposes,suchasmedicalimage,imagedatabaseandsoon.Lossyimageisimportantaswell,becauseofitshighcompressionratio.Inthispaper,weproposeaLosslessimagecompressionSchemeusingalossyimagegeneratedwithPEG-DCTscheme.Ourconceptis,sendaPEG-compressedlossyimageprimary,thensendresidualinformationandreconstructtheoriginalimageusingboththelossyimageandresidualinformation.3-dimensionaladaptivepredictionandanadaptivearithmeticcodingareused,whichfullyusesthestatisticalparameterofdistributionofsymbolsource.Theoptimalnumberofneighborpixelsandlossypixelsusedforpredictionisdiscussed.ThecompressionratioisbetterthanpreviousworkandquiteclosetotheoriginallyLosslessalgorithm.IntroductionTodaytherearemanystudiesonimagecompression,particularlyonlossyandverylowbitratecompression.Forimagedatabase,suchhighcompressionratioisImportantforstorageandalsoforquicktransmission,buttodealwithvariouskindsofusersdemand,Losslessimagetransmissionisindispensable.Inthispaper,weproposeaneffectiveLosslesscompressionalgorithmforgrayimageusinglossycompressedimage.ThelossycompressionschemeusestheJointPhotographicExpertsGroupdiscretecosinetransform(PEG-DCT)algorithmasthelossycodingalgorithm.Firstwesearchthesimilarpairsofpixels(conlexts),accordingtotheirneighborpixels.Forsuchpixelswhichhavecontexts,wepredicttheirvaluesfromthecontextsandtheneighbors.Ontheotherhand,foreachpixelwhichdoesn'thaveitscontextpairs,wecalculatetheedgelevelaccordingtothedifferenceofadjacentpixelvalues.Foreachedgelevelofpixels,wecalculatethepredictivecoefficientsoflinearcombinationundertheleastsquareerrorcriterion.Notonlythepixelswhichhavealreadyprocessedbutalsothepixelsofthelossyimageisusedforprediction.Foreverypixel,thedifferencebetweenpredictedvalueandrealvalueiscalculated,andthedifferenceisconvertedtoanon-negativevaluebeforebeingencoded,accordingtotheirdistribution.Inentropycodingstage,weusethearithmeticcoding.Itismadeadaptive,andinitialerrordistributionisgivenonlybyoneparameter,whichisspecificforeachedgelevel'sstatisticaldistribution.Thepixelsbelongingtothedifferenceedgelevelsareencodedindependently.Experimentalresultsandgoodperformanceareshown.LiketheotherLPL(L0ssyPlusLossless)approaches,ourcompressionratioislessthanthatoforiginallyLosslessscheme,butthedifferenceisslight.Ofallthings,however,usersgetthegreatmeritThattheycanbrowsetheimagebeforetheLosslessdecompression.Manysuchschemeshavebeenproposedintheliterature,butmostofthemtreatthelossyimageanditsLosslessresidualasindependentsymbolsource.OneoftheexceptionsisMemon’salgorithm[6].Weutilizethelossydatathoroughly,andmuchbetterresultisobtained.1.1PixelestimationNormallytheimagedataisscannedalongthescan-linedirection.Infigure1.currentpixelisprocessedpixelsare.OrdinalpixelestimationmethodpredictstheNFL,valueofcurrentpixelusingPI...P4.Thencalculatethepredictionerrore=-.Normallythelinearcombinationisusedforpredictionasfollows,whereTI...T4arethecoefficients.ThisFigure1:Currentpixelistheextrapolativecessedneighborpixels=(1)Usually,thezero-orderentropyofset{e}islowerthanthatofset{z}.Therefore,afterentropycodingschemesuchasHuffmancoding[l]orArithmeticcoding[2]orLempel-Zivcoding[3],datasizeisreduced.Principally,wealsousethelinearcombinationlikeequation(1)forprediction,buttheprocessismoreadaptivethannormalpredictionmethod.Andweusemoreneighborpixels(uptoten),alsousingthepixelsoflossyimageandpredictionerroreisconvertedtoanotherformbeforeencoding.2.1GroupingthepixelsEachimagepixelhasdifferentpropertyundercertaincriterion.Fromapointofimagecompression,groupingsimilarpixelsandencodethemtogethercauseseffectiveresult.Forgroupingthepixels,weusetheQvalue:Q=(2)UsingthisQvalue,weclassifyeachpixelintoseveralgroupsaccordingtotable1Table1:GroupingtableFigure2:(a)Originalimage‘Girl’and(b)JPEGcompressedimage(qua1ityvalue=5)Figure3:(a)ImageofQvalue,(b)ImageofpredictionerrorofsimplepredictionFigure3(a)showstheQvalueand(b)showstheerrorofsimpleprediction.Ascanbeseenfromthem,thevalueQcorrelatecloselywiththepredictionerror.Thereforethepredictioncoefficientsarecalculatedindependentlyineachgroup.2.2ContextsearchTable2showseachgroup’sfinalzero-orderentropyofpredictionresultofimage‘Girl’.Obviously,theuppergroupsaremoredifficulttobecompressedthanthelowergroups.Weusethecontext-basedpredictionmethodtodealwithsuchuppergroups.Theregionwherewesearchthesimilararea(wedenotethis“context〞)isshowninfigure4.Thisisrestrictedwithinalreadyprocessedpixels’area.TheprocedureisDescribedbelow:1.Scantheareaandfindpointsthatsatisfyand(3)2.Withinsuchpoints,findonethatminimizes(4)3.IftheCminissmallerthan12,treatitasthecontextofthecurrentpoint.Otherwise,returnfailure(notfound).2.3Prediction2.3.1PredictionofnormalgroupsForeachgroup,wepredictthevalueofacurrentpixelbylinearcombinationofitsneighbors’pixelvalues.Thecoefficientsarecalculatedbyleastsquareerrormethod.Theneighborpixelsusedforpredictionareshowninfigure5(Pi...Pia).Numberofpixelsarevariable(upto10pixels),andafterwardswewillchooseitoptimumly.Thepriorityisshownbythesuffixnumberinthefigure.Themosteffectivenumberwillbediscussedlater.Table2:GroupV.S.Entropy(image‘Girl’)Figure4:Context-searchregionFigure5:PixelsusedforpredictionHere,someofthelossypixels(RI...&,)gotfromPEG-compressedimageareAlsoused.Usingthesepixels,somethinglikeinterpolativepredictionisachieved.Thispredictioncontributesthereductionofcompressedsize.2.3.2PredictionofcontextUnderthecriterionofcondition(3)and(4).apairofcontexthavethesimilarshapesofheight.Therefore,wepredictthevalue2from{a’,b’,c‘,z‘,a,b,c}(seefigure4).Wealsousetheleastsquareerrorestimation.Themostdifferentpointfrompredictionofnon-contextpixelsis,notonlyusingthevaluesofneighborpixelsbutalsousingtheclosestcontext.Thisschemeiseffectiveforcontinuousedges,becausenearbyanedgethereisasequenceofsimilarlookingpixels.2.4ErrorconversionIfeachpixelhas8bits,thepredictionerrore=-canhavetherealnumberbetween-255and+255(roughly).Afterprediction,eshouldbeexpressedasaninteger.Oneeasywayforconversionis,simplyroundthevalueofftheinteger(calculateLe+0.51)andconsideritasthe2’scomplement8-bitnumber.OurconversionalgorithmistotallybasedonTaniguchi’smethod[5].Afterthisconversion,wecanalsogetthe8-bitnon-negativeintegerE.Firstweobtaintheupperandlowerboundofthegroup(max,min).Then,accordingtothefigure6,converttheactualpixelvalueintointeger.(Inthisfigure,ifactualpixelisequalto“ax’,E=9.)Foreachgroup,wegetthemaximumandminimumpixelvalueandconvertthepredictionerrorrespectively.Thisconversionisreversible.fyougetthepredictedvalueandtheconvertednumberE(andalsoupperandlowerbound),youcanobtaintheactualpixelvaluefromsimilarnumericalline.2.5Adaptivearithmeticcoding2.5.1FittingofthedistributionofEAccordingtotheexperimentalresult,thedistributionofE(figure7(a)),aftererrorconversion,looksveryclosetotherighthalfofaGaussiandistribution:(5)Figure6:Algorithmoferrorconversion(example)Figure7:(a)DistributionofE(image=‘Moon’),(b)FittingofEwherethevarianceis.Asthisdistributionislimitedrighthalf,isequalto,whereNisthenumberofsamples.Thefittingline(estimatedfrequency)ofE’sdistributionisCalculated.Theirgraphsareshowninfigure7(b).YoucanseethatusingthisGaussianmodel,thedistributionofEisapproximatedwell.2.5.2AdaptivearithmeticcodingTheprobabilitydensityfunctionofthisfittingcurvecanbeexpressedbyonlyoneparameter.Thisdistributionisusedtogeneratetheinitialdistributiontableforencoder(andalsoordecoder),insteadofpassingthelargeamountofactualfrequencytable.Forthispurpose,arithmeticcodingisverysuitable.Forthispurpose,ourcoderismadeadaptive,eachsymbolfromthedatastreamrenewsthefrequencytable.Byusingandadaptivearithmeticcoder,evensmallamountofdataiscodedwithhighcodingrate.ExperimentalResultsTogeneratealossyimage,weadopttheJPEGcompressionscheme.ThereasonisthatPEGisthestandardschemeforstillimagecompressionanduserscar1findPEG-toolseasily.Weusethetoolsnamed‘cjpeg’and‘djpeg’,whichistheproductsofIndependentPEGGroup.TocreatethePEGimage,cjpeg-qualityquality-value-optimizefilenameandtodecompressthePEGfile,djpegipeg-fileisinvoked.Theoption’-optimizeperformsoptimizationofentropyencodingparameters.ItusuallymakesthePEGfilealittlesmaller.‘djpeg’hastheoptionTable3:Effectofcontextsearch(N=9,NL4,quality=5)‘blocksmooth’,whichperformscross-blocksmoothing,butfromexperimentalresults,itmadethecompressionratioworse,thereforethisoptionisnotused.Parameterswhicharenecessaryfordecompressionareputtogetherandalsoencodedbyadaptivearithmeticcoder.Onlyseveralpercentcompressionisachieved,butbetterthandoingnothing.3.1EffectofcontextsearchWeusethreetestimages‘Girl’,‘Couple’and‘Moon’,whicharechosenfromSIDBA(StandardImageDataBase).Fortheseimages,thebitrateofcompressingwithandwithoutcontextsearchiscompared.Theresultsareshownintable3.NListhenumberoflossypixelsusedforprediction.NisthenumberofLosslessneighborpixels.3.2EffectofthequalityvalueoncompressionratioPEGprovidesafinetuningfactor(qualityvalue),whichcorrespondstodifferentQualitiesofthecompressedimages.Fortypicalimagedata,alowqualityvaluesuchas20provideshighcompressionwithpoorimagefidelity.Asthequalityvalueincreasesthefidelityimprovesattheexpenseofcompressionratio.Figure8showsthebitrateandqualityvalue.Asthequalityvaluedecreases,thetotalbitratedecreases,too.Laterweusetheimagequalityvalueof5,whichisenoughforunderstandingtheimageroughly(seefigure2(b)).Figure8:BitrateV.S.QualityvalueFigure9:BitrateV.S.NL(N=10,(N=10,-9,images‘Gir1’)quality=5,image=‘Girl’)Table4:Resultofbitrate(quality=5,N=9,NL=4)3.3EffectofusingalossyimageTheremightbeadoubtthatthelossyimagelookssimilarwiththeoriginaltooureyes,butdiffersmuchfromastandpointofpixel-value,thereforeithardlyhelpstheprediction.Figure9showsthebitrateandthenumberoflossypixelsusedforprediction(NL).Fromthisfigure,itisknowneventhepoorimage(quality=5)helpsthecompressionratioconsiderably.ThereasonwhythebitrateincreasesgraduallywhereNLisgreaterthan4is,thatadditionalparameter(coefficients)isnecessaryforeachpixelposition.TheoptimumNLis4.MoreoverweconductedexperimentstofindtheoptimumnumberofN.andobtained9.3.4ComparisonwiththeothermethodsTable4showstheresultofthecompression.Taniguchi’smethod[5]isoriginallyLosslessorientedone,thereforetheresultsareslightlybetterthanours.Inourmethod,NandNLaresetoptimally.MAWmethodisproposedbyMemon[6].ItalsousesalossyPEGimagetoreconstructaLosslessimage.Ourresultisabout0.5bit/pelbetterthanMAW,and0.03to0.12bid\pelworsethanTaniguchi’smethod.AstheMAW’Sresultisgivenonlybyentropy,thedifferenceofperformancefromoursmightbebigger.ConclusionInthispaper,weproposedthealgorithmofLosslessimagecompression.Unlikeotherliteratures,wehavediscussednotonlytheentropyofresidual,buthowtoencodeitefficientlyandthefinalsizeofthecompressedproduct.ThisprovidesanattractiveoptioninapplicationsthathaveneedforquicktransmissionontheonehandandexactReconstructionontheother.Furthermore,usingthislossyimagethoroughly,thetotalbitrateisconsiderablylow.Searchingthecontextsandusingitforpredictionisprovedtowork.Wewillresearchmoreeffectivewayofusingcontext.OuralgorithmisapplicablefororiginallyLosslessorientedone(notusinglossyimage).Insomeimages,wehaveobtainedbetterresultsthanTaniguchi’smethod.Weareproceedingtheinvestigationinthisstandpoint.WeareexaminingifthereisamoresuitablealgorithmforlossycompressionotherthanPEG.AndalsospeculatingthemoreaccurateerrordistributionfittingotherthanGaussianmodel.References[11D.A.Huffman:“Amethodfortheconstructionofminimumredundancycodes〞,ProceedingsofIRE40,pp.411420,1951[2]J.J.Rissanenetal.:“Arithmeticcoding〞,IBMJournalorResearchandDevelopment,23(2),pp.188-193,1976[3]J.ZivandA.Lempel:“AUniversalAlgorithmforSequentialDataCompression〞,IEEETransactionsonInformationTheory,IT-23(3),pp.337-343,May1997[4]PaulG.HowardandJeffereyScottVitter:“NewMethodsforLosslessImageCompressionUsingArithmeticCoding〞,ProceedingsofDataCompressionConference’91,pp.257-266.1991[5]TakayukiTaniguchietal.:“Variable-Length-Code-SelectiveReversiblePredictiveCodingforMulti-LevelImages〞,TheTransactionsoftheInstituteofElectronics,InformationandCommunicationEngineers,Vol.J70-B,pp.654-663,Jun.1987[6]NasirD.Memonetal.:“SimplemethodforenhancingtheperformanceoflossyplusLosslessimagecompressionschemes〞,JournalofElectronicImaging2(3),pp.245-252,Jul.1993[71DmitryA.Novik:“CompressionThroughDecompressionintoBrowseandResidualImages〞,Proceedingsof1993SpaceandEarthScienceDataCompressionWorkshop,NASAConferencePublication3191,pp.7-12,Apr.1993使用自適應(yīng)預(yù)測和自適應(yīng)算術(shù)編碼的有損圖像的無損壓縮摘要多用無損灰度圖像壓縮是必要的，如醫(yī)學(xué)圖像，圖像數(shù)據(jù)庫等。有損圖像很重要，因為其具有較高的壓縮比。在本文中，我們提出了一種無損圖像壓縮，使用JEPG-DCT方案產(chǎn)生的有損圖像格式。我們的主題是，將JEPG壓縮有損形象性，然后把剩余的信息，并使用有損圖像重建原始圖像殘差信息。三維自適應(yīng)預(yù)測和自適應(yīng)算術(shù)編碼的使用，充分利用統(tǒng)計分布源參數(shù)符號。鄰居的最正確數(shù)量像素和像素使用了有損的預(yù)言。問題是更好的比以前的工作相當(dāng)接近原來的無損壓縮算法。引言今天有許多對圖像壓縮的研究，特別是對有損極低比特率壓縮。圖像數(shù)據(jù)庫，這樣高的壓縮比重要的存儲和快速傳輸，但對付各種用戶需求，無損圖像傳輸是必不可少的。在本文中，我們提出了一個有效的無損壓縮算法的灰色使用有損壓縮圖像。有損壓縮方案使用的接頭攝影專家組的離散余弦變換〔JEPG-DCT〕算法為有損編碼算法。首先我們搜索的像素相似〔語境〕，根據(jù)他們的鄰居像素。這樣的背景像素，我們預(yù)測值從上下文和鄰居。另一方面，對于每個像素，沒有上下文對，我們計算了沿水平根據(jù)相鄰像素的值的差異。為每個象素邊緣的水平，我們計算的預(yù)測系數(shù)的線性組合—最小平方誤差準(zhǔn)那么下的國家。不僅像素已經(jīng)處理也有損圖像的像素用于預(yù)測。每個邊緣能級的統(tǒng)計分布。對于每一個像素，得出預(yù)測值和實際值的計算，和是轉(zhuǎn)換到一個非負(fù)的值之前被編碼，根據(jù)他們的分布。在熵編碼階段，我們使用算術(shù)編碼。它是由自適應(yīng)，和初始誤差分布僅由一個參數(shù)給出，這是具體的每個邊緣能級的統(tǒng)計分布。實驗結(jié)果說明，良好的性能。像其他LPL〔有損加上無損〕的方法，我們的壓縮比是小于原來的無損方案，但有輕微的差異。所有的一切，然而，得到用戶的大功他們可以瀏覽圖像無損壓縮解壓縮之前，許多這樣的方案。在文獻(xiàn)中已經(jīng)提出，但他們中的大多數(shù)將有損圖像及其無損剩余作為獨立的符號源。一個例外是內(nèi)存的算法[6]。我們利用有損數(shù)據(jù)徹底，更好的結(jié)果。1.1像素的估計通常的圖像數(shù)據(jù)掃描的討論掃描線方向。圖1。當(dāng)前像素一個加工點。順序像素的預(yù)測NFL使用PI的當(dāng)前像素的值P4。然后，卡爾—方法的預(yù)測誤差e=-通常的線性組合，用于預(yù)測如下，其中Ti~T4是系數(shù)的。此圖1：當(dāng)前像素是外推預(yù)測。處理后的相鄰像素=(1)通常，零階熵集｛ε}低于集{Z}。因此，熵如Huffman編碼[我]或[2]或算術(shù)編碼的編碼方案后基于Lempel-Ziv編碼[3]，減少數(shù)據(jù)大小。正文我們主要是使用的線性組合，如方程〔1〕的預(yù)測，但過程更適應(yīng)比正常的預(yù)測方法。我們使用更多的相鄰像素〔高達(dá)十〕，同時使用有損圖像和預(yù)測誤差E像素轉(zhuǎn)化為另一種形式在編碼之前。2.1個組的像素每個圖像像素具有不同的特性在一定的標(biāo)準(zhǔn)。從一個角度圖像壓縮編碼，分組相似的像素和他們在一起造成有效結(jié)果。分組的像素，我們使用的Q值：Q=(2)使用這個值，我們每個像素分類為幾組，根據(jù)表11、分組表圖2：〔a〕原始圖像的‘女孩’〔b〕JPEG壓縮圖像〔質(zhì)量值=5〕圖3：〔a〕的Q值圖像〔b〕簡單的預(yù)測及預(yù)測誤差圖3〔a〕和〔b〕顯示了Q值誤差可以簡單的預(yù)測?？梢钥匆娝麄?，Q值密切相關(guān)的預(yù)測誤差。因此預(yù)測系數(shù)在每個組獨立計算。2.2上下文搜索表2顯示了每個組的最后的零階熵的圖像預(yù)測結(jié)果“姑娘〞。顯然，上組更難被壓縮比下組。我們使用基于上下文的預(yù)測的方法來處理這種上組。地區(qū)我們搜索類似地區(qū)〔我們表示“語境〞〕所示圖4。這是限制在已經(jīng)處理的像素的面積。該程序是掃描區(qū)找點滿足and(3)在這樣的點，找到一個最小的(4)3、如果最小濃度小于12，把它作為當(dāng)前點的上下文。否那么，返回失敗〔不存在〕。2.3預(yù)測2.3.1預(yù)測正常組我們預(yù)測值的當(dāng)前像素的線性組合鄰居的像素值,系數(shù)計算最小平方誤差法。用預(yù)言鄰域像素圖5所示〔PI……PIA〕的數(shù)量。像素是可變的〔10像素〕，然后我們就選擇最優(yōu)。優(yōu)先顯示在圖中的后綴的數(shù)目,最有效的數(shù)以后討論。表2：集團(tuán)與熵〔圖像的“姑娘〞〕圖4：上下文搜索區(qū)域圖5：用于預(yù)測的像素在這里，一些有損的像素〔RI。..，得到的PEG壓縮圖像〕廣泛地。使用這些像素，到達(dá)像插值預(yù)測。這預(yù)測有助于壓縮尺寸的減少。2.3.2預(yù)測語境條件下的標(biāo)準(zhǔn)〔3〕和〔4〕。一對上下文有相似的形狀的高度。因此，我們預(yù)測值2從{'，B，C，Z，A，B，C}〔參見圖4〕。我們也使用最小平方誤差估計。最大的不同點非背景像素的預(yù)測是，不僅利用相鄰像素的值，但使用最接近的背景。本方案有效的連續(xù)的邊緣，因為附近的邊緣有一個類似的像素序列。2.4轉(zhuǎn)換錯誤如果每個像素有8位，預(yù)測誤差e=-可以有真正的數(shù)255和255之間〔約〕。經(jīng)過預(yù)測，應(yīng)表示為整數(shù)。轉(zhuǎn)化的一個簡單的方法