介紹數(shù)字圖像處理外文翻譯

附錄1外文原文Source:"the21stcenturyliteraturetheappliedundergraduateelectroniccommunicationseriesofpracticalteachingplanTheinformationandcommunicationengineeringspecialtyinEnglishch02_1.PDF120-124Ed:HanDingZhaoJuMin,etcTextA:AnIntroductiontoDigitalImageProcessingIntroductionDigitalimageprocessingremainsachallengingdomainofprogrammingforseveralreasons.Firsttheissueofdigitalimageprocessingappearedrelativelylateincomputerhistory.Ithadtowaitforthearrivalofthefirstgraphicaloperatingsystemstobecomeatruematter.Secondly,digitalimageprocessingrequiresthemostcarefuloptimizationsespeciallyforrealtimeapplications.Comparingimageprocessingandaudioprocessingisagoodwaytofixideas.Letusconsiderthenecessarymemorybandwidthforexaminingthepixelsofa320x240,32bitsbitmap,30timesasecond:10Mo/sec.Nowwiththesamequalitystandard,anaudiostereowaverealtimeprocessingneeds44100(samplespersecond)x2(bytespersampleperchannel)x2(channels)=176Ko/sec,whichis50timesless.Obviouslywewillnotbeabletousethesametechniquesforbothaudioandimagesignalprocessing.Finally,digitalimageprocessingisbydefinitionatwodimensionsdomain;thissomehowcomplicatesthingswhenelaboratingdigitalfilters.Wewillexploresomeoftheexistingmethodsusedtodealwithdigitalimagesstartingbyaverybasicapproachofcolorinterpretation.Asamoreadvancedlevelofinterpretationcomesthematrixconvolutionanddigitalfilters.Finally,wewillhaveanoverviewofsomeapplicationsofimageprocessing.Theaimofthisdocumentistogivethereaderalittleoverviewoftheexistingtechniquesindigitalimageprocessing.Wewillneitherpenetratedeepintotheory,norwillweinthecodingitself;wewillmoreconcentrateonthealgorithmsthemselves,themethods.Anyway,thisdocumentshouldbeusedasasourceofideasonly,andnotasasourceofcode.Asimpleapproachtoimageprocessing(1)Thecolordata:VectorrepresentationBitmapsTheoriginalandbasicwayofrepresentingadigitalcoloredimageinacomputer'smemoryisobviouslyabitmap.Abitmapisconstitutedofrowsofpixels,contractionofthewords“PictureElement”.Eachpixelhasaparticularvaluewhichdeterminesitsappearingcolor.ThisvalueisqualifiedbythreenumbersgivingthedecompositionofthecolorinthethreeprimarycolorsRed,GreenandBlue.Anycolorvisibletohumaneyecanberepresentedthisway.Thedecompositionofacolorinthethreeprimarycolorsisquantifiedbyanumberbetween0and255.Forexample,whitewillbecodedasR=255,G=255,B=255;blackwillbeknownas(R,G,B)=(0,0,0);andsay,brightpinkwillbe:(255,0,255).Inotherwords,animageisanenormoustwo-dimensionalarrayofcolorvalues,pixels,eachofthemcodedon3bytes,representingthethreeprimarycolors.Thisallowstheimagetocontainatotalof256x256x256=16.8milliondifferentcolors.ThistechniqueisalsoknownasRGBencoding,andisspecificallyadaptedtohumanvision.Withcamerasorothermeasureinstrumentswearecapableof“seeing”thousandsofother“colors”,inwhichcasestheRGBencodingisinappropriate.Therangeof0-255wasagreedfortwogoodreasons:Thefirstisthatthehumaneyeisnotsensibleenoughtomakethedifferencebetweenmorethan256levelsofintensity(1/256=0.39%)foracolor.Thatistosay,animagepresentedtoahumanobserverwillnotbeimprovedbyusingmorethan256levelsofgray(256shadesofgraybetweenblackandwhite).Therefore256seemsenoughquality.Thesecondreasonforthevalueof255isobviouslythatitisconvenientforcomputerstorage.Indeedonabyte,whichisthecomputer'smemoryunit,canbecodedupto256values.Asopposedtotheaudiosignalwhichiscodedinthetimedomain,theimagesignaliscodedinatwodimensionalspatialdomain.Therawimagedataismuchmorestraightforwardandeasytoanalyzethanthetemporaldomaindataoftheaudiosignal.Thisiswhywewillbeabletodolotsofstuffandfiltersforimageswithouttransformingthesourcedata,whilethiswouldhavebeentotallyimpossibleforaudiosignal.Thisfirstpartdealswiththesimpleeffectsandfiltersyoucancomputewithouttransformingthesourcedata,justbyanalyzingtherawimagesignalasitis.Thestandarddimensions,alsocalledresolution,forabitmapareabout500rowsby500columns.Thisistheresolutionencounteredinstandardanalogicaltelevisionandstandardcomputerapplications.Youcaneasilycalculatethememoryspaceabitmapofthissizewillrequire.Wehave500x500pixels,eachcodedonthreebytes,thismakes750Ko.Itmightnotseemenormouscomparedtothesizeofharddrives,butifyoumustdealwithanimageinrealtimethenprocessingthingsgettougher.Indeedrenderingimagesfluidlydemandsaminimumof30imagespersecond,therequiredbandwidthof10Mo/secisenormous.WewillseelaterthatthelimitationofdataaccessandtransferinRAMhasacrucialimportanceinimageprocessing,andsometimesithappenstobemuchmoreimportantthanlimitationofCPUcomputing,whichmayseemquitedifferentfromwhatonecanbeusedtoinoptimizationissues.Noticethat,withmoderncompressiontechniquessuchasJPEG2000,thetotalsizeoftheimagecanbeeasilyreducedby50timeswithoutlosingalotofquality,butthisisanothertopic.Vectorrepresentationofcolors

Aswehaveseen,inabitmap,colorsarecodedonthreebytesrepresentingtheirdecompositiononthethreeprimarycolors.Itsoundsobvioustoamathematiciantoimmediatelyinterpretcolorsasvectorsinathree-dimensionspacewhereeachaxisstandsforoneoftheprimarycolors.Thereforewewillbenefitofmostofthegeometricmathematicalconceptstodealwithourcolors,suchasnorms,scalarproduct,projection,rotationordistance.Thiswillbereallyinterestingforsomekindoffilterswewillseesoon.Figure1illustratesthisnewinterpretation:Vectorreprescntatinnofcolorsinathreedimensionsspace①EdgeDetectiFromRWhatwehavesaidbeforewecanquantifythe(2)Immediateapplicationtofilters'difference'①EdgeDetectiFromRWhatwehavesaidbeforewecanquantifythe(2)Immediateapplicationtofilters'difference'betweentwocolorsbycomputingthegeometricdistancebetweenthevectorsrepresentingthosetwocolors.LetsconsidertwocolorsC1=(R1,G1,B1)andC2=(R2,B2,G2),thedistancebetweenthetwocolorsisgivenbytheformula:D(C1,C2)=(R1+R2)2+(G1+G2)2+(B1+B2)2Thisleadsustoourfirstfilter:edgedetection.Theaimofedgedetectionistodeterminetheedgeofshapesinapictureandtobeabletodrawaresult

bitmapwhereedgesareinwhiteonblackbackground(forexample).Theideaisverysimple;wegothroughtheimagepixelbypixelandcomparethecolorofeachpixeltoitsrightneighbor,andtoitsbottomneighbor.Ifoneofthesecomparisonresultsinatoobigdifferencethepixelstudiedispartofanedgeandshouldbeturnedtowhite,otherwiseitiskeptinblack.Thefactthatwecompareeachpixelwithitsbottomandrightneighborcomesfromthefactthatimagesareintwodimensions.Indeedifyouimagineanimagewithonlyalternativehorizontalstripesofredandblue,thealgorithmswouldn'tseetheedgesofthosestripesifitonlycomparedapixeltoitsrightneighbor.Thusthetwocomparisonsforeachpixelarenecessary.Thisalgorithmwastestedonseveralsourceimagesofdifferenttypesanditgivesfairlygoodresults.Itismainlylimitedinspeedbecauseoffrequentmemoryaccess.Thetwosquarerootscanberemovedeasilybysquaringthecomparison;however,thecolorextractionscannotbeimprovedveryeasily.Ifweconsiderthatthelongestoperationsarethegetpixelfunctionandputpixelfunctions,weobtainapolynomialcomplexityof4*N*M,whereNisthenumberofrowsandMthenumberofcolumns.Thisisnotreasonablyfastenoughtobecomputedinrealtime.Fora300x300x32imageIgetabout26transformspersecondonanAthlonXP1600+.Quiteslowindeed.Herearetheresultsofthealgorithmonanexampleimage:algorithm:Noticethatalgorithm:Noticethatthequalityoftheresultsdependsonthesharpnessofthesourceimage.Ifthesourceimageisverysharpedged,theresultwillreachperfection.Howeverifyouhaveaveryblurrysourceyoumightwanttomakeitpassthroughasharpnessfilterfirst,whichwewillstudylater.Anotherremark,youcanalsocompareeachpixelwithitssecondorthirdnearestneighborsontherightandonthebottominsteadofthenearestneighbors.Theedgeswillbethickerbutalsomoreexactdependingonthesourceimage'ssharpness.Finallywewillseelateronthatthereisanotherwaytomakeedgedetectionwithmatrixconvolution.②ColorextractionTheotherimmediateapplicationofpixelcomparisoniscolorextraction.Insteadofcomparingeachpixelwithitsneighbors,wearegoingtocompareitwithagivencolorC1.ThisalgorithmwilltrytodetectalltheobjectsintheimagethatarecoloredwithC1.Thiswasquiteusefulforroboticsforexample.Itenablesyoutosearchonstreamingimagesforaparticularcolor.Youcanthenmakeyourobotgogetaredballforexample.Wewillcallthereferencecolor,theonewearelookingforintheimageC0=(R0,G0,B0).Onceagain,evenifthesquarerootcanbeeasilyremoveditdoesn'treallyimprovethespeedofthealgorithm.WhatreallyslowsdownthewholeloopistheNxMgetpixelaccessestomemoryandputpixel.Thisdeterminesthecomplexityofthisalgorithm:2xNxM,whereNandMarerespectivelythenumbersofrowsandcolumnsinthebitmap.Theeffectivespeedmeasuredonmycomputerisabout40transformspersecondona300x300x32sourcebitmap.3.JPEGimagecompressiontheory(一)JPEGcompressionisdividedintofourstepstoachieve:ColormodeconversionandsamplingRGBcolorsystemisthemostcommonwaysthatcolor.JPEGusesaYCbCrcolorsystem.WanttouseJPEGcompressionmethoddealingwiththebasicfull-colorimages,RGBcolormodetofirstimagedataisconvertedtoYCbCrcolormodeldata.Yrepresentativeofbrightness,CbandCrrepresentsthehue,saturation.Bythefollowingcalculationtobecompletedbydataconversion.Y=0.2990R+0.5870G+0.1140BCb=-0.1687R-0.3313G+0.5000B+128Cr=0.5000R-0.4187G-0.0813B+128ofhumaneyesonthelow-frequencydatathanhigh-frequencydatawithhigherThesensitivity,infact,thehumaneyetochangesinbrightnessthantocolorchangesshouldbemuchmoresensitive,ieYcomponentofthedataismoreimportant.SincetheCbandCrcomponentsisrelativelyunimportantcomponentofthedatacomparison,youcanjusttakepartofthedatatodealwith.Toincreasethecompressionratio.JPEGusuallyhavetwokindsofsamplingmethods:YUV411andYUV422,theyrepresentisthemeaningofY,CbandCrdatasamplingratioofthreecomponents.DCTtransformationThefullnameistheDCT-discretecosinetransform(DiscreteCosineTransform),referstoagroupoflightintensitydataintofrequencydata,inorderthatintensitychangesofcircumstances.Ifthemodificationofhigh-frequencydatado,andthenbacktotheoriginalformofdata,itiscleartherearesomedifferenceswiththeoriginaldata,butthehumaneyeisnoteasytorecognize.Compression,theoriginalimagedataisdividedinto8*8matrixofdataunits.JPEGentireluminanceandchrominanceCbmatrixmatrix,saturationCrmatrixasabasicunitcalledtheMCU.EachMCUcontainsamatrixofnomorethan10.Forexample,theratioofrowsandcolumnsJieWei4:2:2sampling,eachMCUwillcontainfourluminancematrix,amatrixandacolorsaturationmatrix.Whentheimagedataisdividedintoan8*8matrix,youmustalsobesubtractedforeachvalueof128,andthenagenerationofformulaintotheDCTtransformcanbeachievedbyDCTtransformpurposes.Theimagedatavaluemustbereducedby128,becausetheformulaacceptedbytheDCT-figurerangeisbetween-128to+127.QuantizationImagedataisconvertedtothefrequencyfactor,youstillneedtoacceptaquantitativeproceduretoenterthecodingphase.Quantitativephaserequirestwo8*8matrixofdata,oneistodealspecificallywiththebrightnessofthefrequencyfactor,theotheristhefrequencyfactorforthecolorwillbethefrequencycoefficientdividedbythevalueofquantizationmatrixtoobtainthenearestwholenumberwiththequotient,thatiscompletedtoquantify.Whenthefrequencycoefficientsafterquantization,willbetransformedintothefrequencycoefficientsfromthefloating-pointintegerThisfacilitatetheimplementationofthefinalencoding.However,afterquantitativephase,allthedatatoretainonlytheintegerapproximation,alsoonceagainlostsomedatacontent.(4)CodingHuffmanencodingwithoutpatentissues,tobecomethemostcommonlyusedJPEGencoding,HuffmancodingisusuallycarriedoutinacompleteMCU.Coding,eachoftheDCvaluematrixdata63ACvalue,willuseadifferentHuffmancodetables,whilethebrightnessandchromaalsorequireadifferentHuffmancodetables,itneedsatotaloffourcodetables,inordertosuccessfullycompletetheJPEGcoding.DCCodeDCisacolordifferencepulsecodemodulationusingthedifferencecodingmethod,whichisinthesamecomponenttoobtainanimageofeachDCvalueandthedifferencebetweenthepreviousDCvaluetoencode.DCpulsecodeusingthemainreasonforthedifferenceisduetoacontinuoustoneimage,thedifferencemostlysmallerthantheoriginalvalueofthenumberofbitsneededtoencodethedifferencewillbemorethantheoriginalvalueofthenumberofbitsneededtoencodetheless.Forexample,amarginof5,anditsbinaryrepresentationofavalueof101,ifthedifferenceis-5,thenthefirstchangedtoapositiveinteger5,andthenconvertedintoits1'scomplementbinarynumbercanbe.Theso-calledone'scomplementnumber,thatis,ifthevalueis0foreachBit,thenchangedto1;Bitis1,itbecomes0.Differencebetweenthefiveshouldretainthemedian3,thefollowingtablethatliststhedifferencebetweentheBittoberetainedandthedifferencebetweenthenumberofcontentcontrols.Inthemarginofthemarginfront-endaddsomeadditionalvalueHoffmancode,suchasthebrightnessdifferenceof5(101)ofthemedianofthree,thentheHuffmancodevalueshouldbe100,thetwoconnectedtogethershallbe100101.ThefollowingtwotablesarethebrightnessandchromaDCdifferenceencodingtable.Accordingtothesetwoformscontent,youcanaddthedifferencefortheDCvalueHuffmancodetocompletetheDCcoding.4.ConclusionsDigitalimageprocessingisfarfrombeingasimpletransposeofaudiosignalprinciplestoatwodimensionsspace.Imagesignalhasitsparticularproperties,andthereforewehavetodealwithitinaspecificway.TheFastFourierTransform,forexample,whichwassuchapracticaltoolinaudioprocessing,becomesuselessinimageprocessing.Oppositely,digitalfiltersareeasiertocreatedirectly,withoutanysignaltransforms,inimageprocessing.Digitalimageprocessinghasbecomeavastdomainofmodernsignaltechnologies.Itsapplicationspassfarbeyondsimpleaestheticalconsiderations,andtheyincludemedicalimagery,televisionandmultimediasignals,security,portabledigitaldevices,videocompression,andevendigitalmovies.Wehavebeenflyingoversomeelementarynotionsinimageprocessingbutthereisyetalotmoretoexplore.Ifyouarebeginninginthistopic,Ihopethispaperwillhavegivenyouthetasteandthemotivationtocarryon.附錄2外文翻譯文獻出處：《21世紀全國應(yīng)用型本科電子通信系列實用規(guī)劃教材》之《信息與通信工程專業(yè)英語》ch02_1.pdf120-124頁主編：韓定定、趙菊敏等正文：介紹數(shù)字圖像處理.導(dǎo)言有幾個原因使數(shù)字圖像處理仍然是一個具有挑戰(zhàn)性的領(lǐng)域。首先，出現(xiàn)數(shù)字圖像處理問題相對在計算機歷史的后期。從第一個圖形化操作系統(tǒng)的到來，才成為了真正的問題。其次，數(shù)字圖像處理要求的是對實時應(yīng)用慎重優(yōu)化。比較圖像處理和音頻處理是一個很好的解決的想法。讓我們考慮審查為320x240，32位位圖，30次：10鉬/秒的像素所需的內(nèi)存帶寬。同樣的質(zhì)量標準，音頻立體聲波實時處理現(xiàn)在需要44100（每秒采樣）義2（字節(jié)每通道采樣）義2（通道）=176Ko/sec,這是小50倍。很明顯，我們將無法使用音頻和圖像信號處理相同的技術(shù)。最后，數(shù)字圖像處理是兩個層面的定義域，在擬定的數(shù)字濾波器中是有點復(fù)雜的事情。我們將探討一些現(xiàn)存的方法去處理與解釋數(shù)碼圖像出發(fā)與色彩非常基本的觀念。來作為解釋更先進水平的矩陣卷積和數(shù)字濾波器。最后，我們將給出一些應(yīng)用與圖像處理的概述。本文件的目的是給出讀者現(xiàn)有的數(shù)字圖像處理技術(shù)的概述。我們既不會滲透更深的理論，也不會自己編碼，我們將更加專注于自身的算法，該方法。無論如何，本文件應(yīng)作為一種有思想的來源，而不是代碼的來源。.一個簡單的圖像處理方法（1）色彩數(shù)據(jù)：矢量表示①位圖該數(shù)字代表的顏色在計算機的內(nèi)存映像原件和根本途徑中顯然是一個位圖。構(gòu)成位圖的像素行，簡稱為“像素”。每個像素都有一個特定的數(shù)值，決定了它出現(xiàn)的顏色。此值是由三基色分解為紅，綠，藍。任何肉眼可見的顏色都可以用這種方式的代表。一個色在三基色分解量化數(shù)量介于0和255之間。例如，白將被編碼為R=255,G=255,B=255;黑將是為（R,G和B）=（0,0,0）這樣可得之，明亮的粉紅色將是：（255,0,255）。換言之，一個圖像是一個巨大的兩色值的二維數(shù)組，像素，每個編碼的3個字節(jié)，代表三基色。這使得圖像包含256義256X256=16.8萬種不同的顏色總數(shù)。這種技術(shù)也被稱為RGB編碼，并特別適合人類視覺。用相機或其他措施的文書，我們有能力“看見”其他“千色”，在這種情況下，RGB編碼是不適當?shù)?。把范圍定?-255有兩個很好的理由：第一，人眼不足夠的理性，使超過256個之間的強度（1/256=0.3%的顏色）水平的差異。這就是說，提交給人類觀察者的圖像，將無法通過使用改進的超過256個灰色（256深淺不同的層次灰黑色和白色之間）。因此，256有足夠的優(yōu)點。值為255的第二個原因是很明顯，它是計算機存儲方便。事實上，在一個字節(jié)，它是計算機的內(nèi)存單元，可以編碼256值。相對于音頻信號，在時域編碼中圖像信號被編碼在一個二維空間域。原始圖像數(shù)據(jù)更加簡單和易于分析比音頻信號的時域數(shù)據(jù)。這就是為什么我們將能夠不改變源數(shù)據(jù)來做圖像和過濾器更多的東西，而音頻信號將是不可能完成的。這第一部分處理簡單的效果和過濾器分涉您可以通過計算不改變源數(shù)據(jù)，只是通過分析原始圖像信號來實現(xiàn)。標準尺寸，也稱為分辨率，是一個位約500行500列。分辨率運用于標準的電視和標準的計算機應(yīng)用中。您可以輕松地計算出這樣大小的內(nèi)存空間需要多少位圖。我們有500X500像素，每3個字節(jié)編碼，這使得750高。它看起來沒有龐大的規(guī)模比較硬碟，但如果你要用實時圖像處理事情就更加嚴格。事實上，渲染圖像流暢的要求為至少30個圖像，10鉬/秒的巨大需要的帶寬。我們會在后面看到的數(shù)據(jù)的限制準入和移交內(nèi)存在圖像處理極為重要，有時它正好是遠遠超過CPU的計算，限制重要的工作似乎頗為可以用來在一個不同的是什么優(yōu)化問題。請注意，用現(xiàn)代的壓縮技術(shù)，如JPEG2000,總大小的圖像可以很容易地減少50倍而不失去很多的質(zhì)量，但是這是另一個話題。②代表性顏色矢量正如我們所見,在一個位圖,顏色編碼在三個字節(jié)表示他們對三原色分解。這聽起來是顯而易見的，一個數(shù)學(xué)家解釋為顏色在三維空間向量各軸代表原色。因此，我們將利用數(shù)學(xué)概念的幾何來處理大部分的顏色，如規(guī)范，標產(chǎn)品，投影，旋轉(zhuǎn)或距離。這將是很有趣的過濾器的一些實物，我們將很快看到。圖1說明了這種新的解釋：GreenVectorrepresciitaiiotiofcolorsinathreedimensionsspaceCyan=（（U55,255）Ycl]ow=（255,255.0）GreenVectorrepresciitaiiotiofcolorsinathreedimensionsspaceCyan=（（U55,255）Ycl]ow=（255,255.0）圖1⑵即時過濾器的應(yīng)用①邊緣檢測我們已經(jīng)說過，我們可以通過兩種顏色之間量化計算，代表這兩種顏色向量之間的幾何距離，讓我們考慮兩種顏色C1=（R1，G1，B1）和C2=（R2,B2,G2），在兩種顏色之間的距離是由下列公式：D(C1,C2)=(R1+R2)2+(G1+G2)2+(B1+B2)2這使我們對第一個過濾器：邊緣檢測算法。目的是確定邊緣檢測圖像邊緣的形狀，可以得出圖中白色邊在黑色的背景（例如）。這個想法很簡單，我們一起經(jīng)歷了由像素圖像象素的顏色,并比對鄰居的每一個像素,在它的底部的鄰居。如果這些比較結(jié)果存在太大差異，像素邊緣研究的一部分是要變白的,否則它保存在黑色中。事實上,我們比較了每一個象素與它的底部和正確的鄰居來源于的圖像是在兩個維度上。事實上，如果你選擇紅色和藍色，算法不會看到這些條紋的邊緣圖像的水平條紋，如果只是一個像素相比，其是右邊的鄰居。因此，對每個像素的兩個比較是必要的。該算法對源圖像檢測的幾種不同類型中存在的問題，提出了比較好的結(jié)果。它主要局限在速度，因為頻繁的內(nèi)存的訪問。在兩個可以很容易被清除的平方根中比較，顏色提取物不能很容易得到改善。如果我們認為是世界上最長的操作功能，像素函數(shù)獲得了4*N*M個復(fù)雜多項式，其中N是行和M數(shù)列的數(shù)目。這是相當快的，可以用實時計算。對于300X300X32的圖像我得到的約26轉(zhuǎn)換每秒AthlonXP1600+。相當緩慢。這里是一個例子形象，該算法的結(jié)果：WW圖像存在非原，首先，你要經(jīng)過銳利之后濾過。你還可以比較每個像素和它的第二或第三最近鄰居在右邊底部代替為最近的鄰居。邊緣厚而且會更精確取決于源圖像的鋒芒。最后，我們將會看到,還有另一種方法使邊緣檢測與矩陣回旋。②顏色提取其他比較像素的應(yīng)用是直接從顏色提取的。相反,每一個象