`
`Robust transmission of compressed images over
`noisy Gaussianchannels
`
`Conference Paper · June 1995
`
`DOI: 10.1109/ICASSP.1995.479956 · Source: IEEE Xplore
`
`CITATION
`1
`
`6 authors, including:
`
`READS
`5
`
`Robert L. Stevenson
`University of Notre Dame
`
`Yih-Fang Huang
`University of Notre Dame
`
`153 PUBLICATIONS 3,784 CITATIONS
`
`171 PUBLICATIONS 3,680 CITATIONS
`
`SEE PROFILE
`
`SEE PROFILE
`
`Some of the authors of this publication are also working on these related projects:
`
`Efficient training for deep learning View project
`
`Adaptive distribution estimation View project
`
`All content following this page was uploaded by Robert L. Stevenson on 03 September 2014.
`
`The user has requested enhancement of the downloaded file. All in-text references underlined in blue are added to the original document
`
`and are linked to publications on ResearchGate, letting you access and read them immediately.
`
`IPR2016-01179 Ex. 2012
`Page 1 of 5
`
`
`
`ROBUSTTRANSMISSIONOFCOMPRESSEDIMAGESOVERNOISY
`GAUSSIANCHANNELS
`ThomasP.O'Rourke,RobertL.Stevenson,Yih-FangHuang,
`LanceC.PerezandDanielJ.CostelloJr.
`LaboratoryforImageandSignalAnalysis
`DepartmentofElectricalEnginering
`UniversityofNotreDame
`NotreDame,INUSA
`ExperimentalresultsareshowninSection toillustrate
`ABSTRACT
`Manyimagecommunicationsystemshaveconstraintson
`theconceptsinvolved.Theresultsofsimulationexperi-
`bandwidth,powerandtimewhichprohibittransmissionof
`mentsalsoshowtheaverageperformanceoftheproposed
`uncompressedrawimagedata.Compressedimageformats,
`systemforvaryingnoiselevels.
`however,areextremelysensitivetobiterrorswhichcan
`seriouslydegradethequalityoftheimageatthereceiver.
`.SYSTEMSUMMARY
`Anewlist-basediterativetrellisdecoderisproposedwhich
`acceptsfeedbackfromapost-processorwhichcandetect
`Ablockdiagramoftheproposedimagecommunicationsys-
`channelerrorsinthereconstructedimage.Experimental
`temisshowninFigure .
`resultsareshownwhichindicatethenewdecoderprovides
`signi(cid:12)cantimprovementoverthestandardViterbidecoder.
`. .Transmitter
` .INTRODUCTION
`Theinputimageiiscompressedbythesourceencoderus-
`ingtheJPEGstillimagecompressionstandard[ ].JPEG's
`Thesensitivityofthecompressedimagerepresentationto
`extendedsequentialmodeofoperationisusedwithcustom
`biterrorsrequiresapplicationofachannelcodebeforetrans-
`quantizationtables,optimizedHu(cid:11)mancodingtables,and
`missionovernoisychannels.Topreventtheuncontrolled
`restartmarkersaftereachrowofblocks.Therestartmark-
`degradationcausedbyachannelerror,anerrorcontrol-
`erslimitthein(cid:13)uenceofachannelerrortoasinglerowof
`lingchannelcodeisappliedtothecompressedrepresenta-
`blocks.Thecompressedrepresentationbisencodedforthe
`tionbeforetransmission.Thecostoftheadditionalbits
`noisychannelusingarate /convolutionalcodewithcon-
`forredundancyinthechannelcodeispaidforbyanin-
`straintlength[].Thebitsequenceb(cid:3)isthentransmitted
`creasedcompressionratiowhichresultsinadditionalcon-
`overthenoisychannelusingBPSKmodulation.
`trolledquantizationerror.
`Althoughthechannelcodegreatlyreducesthenumber
`..Receiver
`oferrorsinthecompressedimagerepresentation,asingle
`errorcouldstillproduceseveredegradationinthequalityof
`AniterativedecoderbasedonasoftdecisionViterbitrellis
`decoderinterpretsthenoisyreceivedbit-stream^b(cid:3).The
`thereceivedimage.Thepost-processingmethodforreduc-
`ingthevisibilityofquantizationerrorspresentedin[ ,]
`(cid:12)rstiterationdecodesthestandardsoftdecisiontrellisto
`makesuseoftheHuberMarkovrandom(cid:12)eld(HMRF)im-
`obtainthemaximumlikelihoodsequence^bgiventhere-
`agemodel.Therobustimagecommunicationsystempro-
`ceivedchannelsymbols.However,itisalsoknownthat^bis
`posedhereusesthisimagemodeltodetecterrorsinthe
`aJPEGcompressedimagerepresentation.Sinceacorrect
`compressedimagerepresentationandfeedsthiserrorinfor-
`decodingoftheJPEGheaderinformationiscriticaltothe
`mationbacktothechanneldecoderforasecondpassat
`correctreconstructionoftheimage,theseconditeration
`decodingthechannelsymbols.Afterchannelerrorshave
`redecodesthesectionofthetrelliscontainingtheJPEG
`beencorrected,theimageispost-processedtoreducethe
`header.Theheadersyntaxde(cid:12)nedbytheJPEGstandard
`visibilityofthequantizationerror.Unlikeotheralgorithms,
`determinesthevalueofmanybitsintheheaderandallows
`thissystemcoordinateschannelerrorrecoverywithquanti-
`detectionofincorrectheaderinformation.Theknownbits
`zationerrorreduction.Anewiterativechanneldecoder
`reducethenumberofpathsthroughthetrellisanddecrease
`acceptserrorfeedbackfromthenowdual-purposepost-
`theprobabilityofdecodinganincorrectpath.Thisisvery
`processor.
`InSection,amoredetailedsummaryofthe
`similartothepinnedstatedecoderdescribedin[].
`proposedimagecommunicationsystemwillbepresented.
`Thethirditerationconsiderstheheadertobeknown
`correctlyandredecodessectionsofthetrelliscorresponding
`ThisworkwassupportedinpartbyNASALewisResearch
`toentropycodedimagedatawhichhavebeensignaledby
`CenterundercontractNASA-NAG - .ToappearinProc.
`ICASSP- ,(Detroit,MI),May- ,
`thepost-processoraspossiblesitesforerrorevents.The
`
`IPR2016-01179 Ex. 2012
`Page 2 of 5
`
`
`
`Source
`Image
`
`i
`
`i^
`
`Output
`Image
`
`b
`
`Source
`Coder
`Q[A .]
`
`Channel
`Coder
`
`b*
`
`Modulator
`
`s^
`
`Source
`Decoder
`R[.]
`
`b^
`
`Channel
`Decoder
`
`b*^
`
`Demodulator
`
`c
`
`c^
`
`Noisy
`Channel
`
`Post-
`Processing
`F(.)
`
`Figure :Proposedimagecommunicationsystem
`successofthethirditerationdependsontheabilityofthe
`post-processortodetecterroreventsinthereconstructed
`image.TheerrorsaredetectedusingtheHuber-Markov
`random(cid:12)eld(HMRF)imagemodel.See[ ,]formore
`informationontheHMRFimagemodel.TheHMRFmodel
`ischaracterizedbyaspecialformoftheGibbsdistribution
`Pr(x)= Zexpf(cid:0) (cid:21)XcC(cid:26)T(dtcx)g
`where(cid:21)isascalarconstantthatisgreaterthanzero,dc
`isacollectionoflinearoperatorsandthefunction(cid:26)T((cid:1))is
`givenby(cid:26)T(u)=(cid:26)u;
`juj(cid:20)T;
`T+T(juj(cid:0)T);
`juj>T:
`Thismodelisusedtodetecterrorsinaregionoftheimage
`byestimatingtheprobabilityofthatregion.Regionswhich
`aregreatlya(cid:11)ectedbychannelerrorswillhavealargevalue
`fortheexponenttermPcC(cid:26)T(dtcx)andtheprobability
`measurefortheseregionswillbeverylow.
`Anerroreventproducesthreedi(cid:11)erenttypesofarti-
`factsinthereconstructedimage.AmissedEnd-of-Block
`Figure:Originalairportimage,(cid:2).
`codewillcauseanincorrectnumberofblocksforapartic-
`ularrow.Whileanincorrectnumberofblocksindicates
`anerrorhasoccurred,this(cid:12)rsttypeoferrordoesnotpro-
`Informationaboutpossibleerrorlocationsisfedbackto
`videinformationonwhereintherowtheerroroccurred.
`thetrellisdecoderforreconsideration.Boundariesbetween
`Second,anerrorintheDCtermwillpropagateuntilthe
`rowsandindividualblockswhichhaveprobabilitiesbelow
`nextrestartmarkerattheendoftherow.Thiserrorcan
`aparticularerrordetectionthresholdareconsideredpossi-
`bedetectedbycalculatingtheprobabilityfromtheimage
`bleerrorregionsandthecorrespondingsectionsofthebit-
`modelfortheboundaryareabetweenthecurrentrowand
`streamaremarked.Topreventfalsealarms,thelocationsof
`thepreviousrow.ThethirdtypeoferroroccursintheAC
`thethree(cid:2)imageblockswiththelowestprobabilityare
`coe(cid:14)cientsandoftencausesasingle(cid:2)blocktodi(cid:11)er
`giventothedecoderassideinformation.Additionally,the
`greatlyfromtheblocksexpectedbytheimagemodel.This
`errordetectionthresholdwhichiscalculatedforthepartic-
`errorisdetectedbycalculatingtheimagemodeloneach
`ularimageisalsogiventothedecoderassideinformation.
`(cid:2)blockandismosteasilydetectedwhenlargehighfre-
`Thissmallamountofsideinformationcanbeincludedin
`quencycomponentsarepresent.Thisthirdtypeoferroris
`theheaderwithadditionalredundancyforerrorprotection.
`mostusefulsincethelocationoftheerrorwithintherow
`canbecalculated.The(cid:12)rstandlastbitsoftheroware
`TheViterbidecodermakesabranchdecisionateach
`indicatedbyrestartmarkers.Theregionofdoubtiscal-
`statetoselecttheincomingpathwiththelowestweight.
`culatedas(cid:6) %ofthebitsintherowandiscenteredat
`Whenthepost-processorquestionsthedecodingofthetrel-
`theestimatedpositionofthelow-probabilityblockinthe
`lis,thecon(cid:12)dencewithwhicheachbranchdecisionismade
`bit-stream.Sinceerroreventsfromthechanneldecoder
`isenteredintoalistforeachstatealongthemostlikelypath
`canproduceaburstoferrors,acombinationofthesethree
`intheregionofdoubt.Thislistissortedwiththeleast
`typesofartifactsareoftenfoundtogether.
`con(cid:12)dentdecisionatthetop.Thebranchdecisionwith
`
`IPR2016-01179 Ex. 2012
`Page 3 of 5
`
`
`
`Figure :AirportimagecompressedbyJPEGto . bpp,
`noerrors.
`leastcon(cid:12)denceisoverturnedandthenewpaththrough
`thetrellisisdecoded,uncompressed,andsenttothepost-
`processor.Theprocesscontinuesoverturningbranchdeci-
`sionsinthesortedlistuntilthepost-processordoesnotsig-
`nalanerrorinthissectionortheendofthelistisreached.
`Onlyonebranchdecisionisoverturnedatatimesinceit
`isassumedtheregionofdoubtcontainsonlyasingleerror
`event.Topreventerroneousredecodingduetofalsealarms
`signaledbythepost-processor,thelengthofthelistislim-
`itedtocontainonlybranchdecisionswhichweremadewith
`con(cid:12)dencelessthanaparticularthresholdvalue.
` .RESULTS
`Experimentswererunusingtheairportimageshownin
`Figure.Thisimagewascompressedto : bpp,seeFig-
`ure .ImageSNRisusedheretomeasureimagequality.
`Althoughsubjectiveimageevaluationismoremeaningful,
`anobjectivequalitymeasurewasneededtoillustrateper-
`formanceaveragedoverseveraltrials.Thecompressionre-
`ducestheimageSNRto . dB.ChannelSNR(Ep=N )
`isexpressedindBwhereEpistheenergyperpixel.Since
`thecompressedimagehas : bpp,thisisequaltothe
`morecommonEb=N whereEbistheenergyperinforma-
`tionbit.UsingEp=N willallowcomparisonofsystems
`withdi(cid:11)erentcompressionratios.
`Theimportanceofcorrectdecodingoftheimageheader
`isshowninFigure.Anerrorinthequantizationtable
`afterstandardViterbidecodinghasseverelydegradedthe
`image(SNR .dB).Thiserroriscorrectedinthesecond
`iteration.Theresultingimage(SNR .dB)contains
`onlyonesmallerrorwhichisnotverynoticeableandnot
`detectedinthethirditeration.
`Sincetheimageheaderconsistsofarelativelysmall
`
`Figure:Exampleofquantizationtableerror,after(cid:12)rst
`iteration,ChannelSNR . dB.
`numberofbits,mostoftheerroreventsappearinthelarger
`entropycodedimagebody.Thefeatureswhichmakean
`errorhighlyvisiblecanbeseeninFigurewhichshows
`anexamplewithtwoerroreventsafterthe(cid:12)rstiteration
`(SNR . dB).Thee(cid:11)ectofeachchannelerrorislimited
`toasinglerowbytherestartmarkers.The(cid:12)rsterrorevent
`causedanextrablocktobeinsertedshiftingtherowtothe
`right.
`Intheseconderrorevent,amissingEnd-of-Block
`codecausedthenextblocktobetreatedashighfrequency
`informationwhichshiftedtheremainderoftherowtothe
`left.ADCerrorisalsopropagatedthroughtherestofboth
`rows.Bothoftheseerroreventsarecorrectedinthethird
`iterationresultinginanimageidenticaltotheerrorfree
`imageshowninFigure .Thequantizationerrorreduction
`bythepost-processorisnotshownhere.
`Whiletheaboveexamplesshowverygooderrorcor-
`rection,theactualperformancewillvarydependingonthe
`particularrealizationofthenoise.Di(cid:11)erentnoisesequences
`ofequalpowercanhaveverydi(cid:11)erente(cid:11)ectsontherecon-
`structedimages.Figureshowstheaverageperformance
`ofthesystemunderconsideration. trialswerecon-
`ductedforeachoftheninechannelSNRlevels.Asex-
`pected,theimageSNRincreasesasthechannelSNRin-
`creases.Thequantizationnoiseduetocompressionlim-
`itstheperformanceathighchannelSNR.Performanceaf-
`terstandardViterbidecoding,correspondingtothe(cid:12)rst
`iteration,isshownwiththesolidline.Thedottedline
`showsperformanceaftertheseconditerationhascorrected
`headererrors.Thedashedlineshowsperformanceafterthe
`thirditerationhascorrectederrorsintheimagebody.Im-
`ageswhichareseverelydegradedbyheadererrorsimprove
`tremendouslywhentheerrorintheheaderiscorrected.
`Althoughmoreimageshaveerrorsintheimagebody,the
`degradationswhicharecorrectedarelesssevere.
`
`IPR2016-01179 Ex. 2012
`Page 4 of 5
`
`
`
`23
`
`22
`
`21
`
`20
`
`19
`
`18
`
`17
`
`Image SNR (dB)
`
`4.2
`4.1
`4
`3.9
`Channel SNR Ep/N0 (dB)
`
`Figure:ImageSNRvs.ChannelSNR
`.CONCLUSION
`Thenewiterativetrellisdecoderisabletoovercomechan-
`nelnoiseusingknowledgeofcompressedimagesyntaxand
`theHMRFimagemodel.Theresultsarescalabletodi(cid:11)er-
`entdegreesofquantizationandcanbeextendedtoother
`compressiontechniques.Additionalerrorprotectionispos-
`siblebyusingalongerconstraintconvolutionalcodeatthe
`expenseofadditionalreceivercomplexity.
`.REFERENCES
`[ ]R.L.Stevenson,\Reductionofcodingartifactsin
`transformimagecoding,"inProc.ICASSP- ,(Min-
`neapolis,MN),pp.V: { ,Apr. .
`[]T.P.O'RourkeandR.L.Stevenson,\Improvedim-
`agedecompressionforreducedtransformcodingarti-
`facts,"inProc.SPIEImageandVideoProcessingII,
`vol. ,(SanJose,CA,Feb.- , ),pp. - ,
` .
`[ ]W.B.PennebakerandJ.L.Mitchell,JPEG:StillIm-
`ageDataCompressionStandard.NewYork:VanNos-
`trandReinhold, .
`[]S.LinandD.J.Costello,Jr.,ErrorControlCoding:
`FundamentalsandApplications,EnglewoodCli(cid:11)s,NJ:
`Prentice-Hall, .
`Figure:Examplewitherroreventsinentropycoded
`data,after(cid:12)rstiteration,ChannelSNR . dB.
`[]O.CollinsandM.Hizlan,\DeterminateStateCon-
`volutionalCodes,"IEEETrans.onCommunications,
`vol. ,pp. - ,Dec. .
`
`16
`3.6
`
`3.7
`
`3.8
`
`4.3
`
`4.4
`
`4.5
`
`
`
`View publication statsView publication stats
`
`IPR2016-01179 Ex. 2012
`Page 5 of 5
`
`