数字图像增强外文文献附翻译

DigitalImageProcessingandEdgeDetectionDigitalImageProcessingInterestindigitalimageprocessingmethodsstemsfromtwoprincipalapplica-tionareas:improvementofpictorialinformationforhumaninterpretation;andprocessingofimagedataforstorag,transmission,andrepresentationforau-se.Animagemaybedefinedasatwo-dimensionalfunction,f(x,y),wherexandyarespatial(plane)coordinate,andtheamplitudeoffatanypairofcoordinates(x,y)iscalledtheintensityorgryleveloftheimageatthatpoint.henx,y,andtheamplitudevaluesoffareallfinit,discretequantitie,wecalltheimageadigitalimage.hefieldofdigitalimageprocessingreferstoprocessingdigitalimagesbymeansofadigitalcompute.Notethatadigitalimageiscomdfaerf,hfhsarnvalu.heseelementsarereferredtoaspictureelements,imageelements,pels,andpiels.Pielisthetermmostwidelyusedtodenotetheelementsofadigitalimag.Visionisthemostadvancedofoursenses,soitisnotsurprisingthatimagesplaythesinglemostimportantroleinhumanperception.However,unlikehumans,whoarelimitedtothevisualbandoftheelectromagnetic(EM)spec-trum,imagingmachinescoveralmosttheentireEMspectrum,rangingfromgammatoradiowaves.Theycanoperateonimagesgeneratedbysourcesthathumansarenotaccustomedtoassociatingwithimages.Theseincludeultra-sound,electronmicroscopy,andcomputer-generatedimages.Thus,digitalimageprocessingencompassesawideandvariedfieldofapplications.Thereisnogeneralagreementamongauthorsregardingwhereimageprocessingstopsandotherrelatedareas,suchasimageanalysisandcomputervi-sion,start.Sometimesadistinctionismadebydefiningimageprocessingasadisciplineinwhichboththeinputandoutputofaprocessareimages.Webelievethistobealimitingandsomewhatartificialboundary.Forexample,underthisdefinition,eventhetrivialtaskofcomputingtheaverageintensityofanimage(whichyieldsasinglenumber)wouldnotbeconsideredanimageprocessingoperation.Ontheotherhand,therearefieldssuchascomputervisionwhoseultimategoalistousecomputerstoemulatehumanvision,includinglearningandbeingabletomakeinferencesandtakeactionsbasedonvisualinputs.Thisareaitselfisabranchofartificialintelligence(AI)whoseobjectiveistoemulatehumanintelligence.ThefieldofAIisinitsearlieststagesofinfancyintermsofdevelopment,withprogresshavingbeenmuchslowerthanoriginallyanticipated.Theareaofimageanalysis(alsocalledimageunderstanding)isinbe-tweenimageprocessingandcomputervision.Therearenoclearcutboundariesinthecontinuumfromimageprocessingatoneendtocomputervisionattheother.However,oneusefulparadigmistoconsiderthreetypesofcomputerizedprocessesinthiscontinuum:low-,mid-,andhighlevelprocesses.Low-levelprocessesinvolveprimitiveopera-tionssuchasimagepreprocessingtoreducenoise,contrastenhancement,andimagesharpening.Alow-levelprocessischaracterizedbythefactthatbothitsinputsandoutputsareimages.Mid-levelprocessingonimagesinvolvestaskssuchassegmentation(partitioninganimageintoregionsorobjects),descriptionofthoseobjectstoreducethemtoaformsuitableforcomputerprocessing,andclassification(recognition)ofindividualobjects.Amidlevelprocessischaracterizedbythefactthatitsinputsgenerallyareimages,butitsoutputsareattributesextractedfromthoseimages(e.g.,edges,contours,andtheidentityofindividualobjects).Finally,higherlevelprocessinginvolves“makingsense”ofanensembleofrecognizedobjects,asinimageanalysis,and,atthefarendofthecontinuum,performingthecognitivefunctionsnormallyassociatedwithvision.Basedontheprecedingcomments,weseethatalogicalplaceofoverlapbetweenimageprocessingandimageanalysisistheareaofrecognitionofindividualregionsorobjectsinanimage.Thus,whatwecallinthisbookdigitalimageprocessingencompassesprocesseswhoseinputsandoutputsareimagesand,inaddition,encompassesprocessesthatextractattributesfromimages,uptoandincludingtherecognitionofindividualobjects.Asasimpleillustrationtoclarifytheseconcepts,considertheareaofautomatedanalysisoftext.Theprocessesofacquiringanimageoftheareacontainingthetext,preprocessingthatimage,extracting(segmenting)theindividualcharacters,describingthecharactersinaformsuitableforcomputerprocessing,andrecognizingthoseindividualcharactersareinthescopeofwhatwecalldigitalimageprocessinginthisbook.Makingsenseofthecontentofthepagemaybeviewedasbeinginthedomainofimageanalysisandevencomputervision,dependingonthelevelofcomplexityimpliedbythestatement“makingsense.”Aswillbecomeevidentshortly,digitalimageprocessing,aswehavedefinedit,isusedsuccessfullyinabroadrangeofareasofexceptionalsocialandeconomicvalue.Theareasofapplicationofdigitalimageprocessingaresovariedthatsomeformoforganizationisdesirableinattemptingtocapturethebreadthofthisfield.Oneofthesimplestwaystodevelopabasicunderstandingoftheextentofimageprocessingapplicationsistocategorizeimagesaccordingtotheirsource(e.g.,visual,X-ray,andsoon).Theprincipalenergysourceforimagesinusetodayistheelectromagneticenergyspectrum.Otherimportantsourcesofenergyincludeacoustic,ultrasonic,andelectronic(intheformofelectronbeamsusedinelectronmicroscopy).Syntheticimages,usedformodelingandvisualization,aregeneratedbycomputer.Inthissectionwediscussbrieflyhowimagesaregeneratedinthesevariouscategoriesandtheareasinwhichtheyareapplied.ImagesbasedonradiationfromtheEMspectrumarethemostfamiliar,es-peciallyimagesintheX-rayandvisualbandsofthespectrum.Electromagnet-icwavescanbeconceptualizedaspropagatingsinusoidalwavesofvaryingwavelengths,ortheycanbethoughtofasastreamofmasslessparticles,eachtravelinginawavelikepatternandmovingatthespeedoflight.Eachmasslessparticlecontainsacertainamount(orbundle)ofenergy.Eachbundleofenergyiscalledaphoton.Ifspectralbandsaregroupedaccordingtoenergyperphoton,weobtainthespectrumshowninfig.below,rangingfromgammarays(highestenergy)atoneendtoradiowaves(lowestenergy)attheother.ThebandsareshownshadedtoconveythefactthatbandsoftheEMspectrumarenotdistinctbutrathertransitionsmoothlyfromonetotheother.Imageacquisitionisthefirstprocess.Notethatacquisitioncouldbeassimpleasbeinggivenanimagethatisalreadyindigitalform.Generally,theimageacquisitionstageinvolvespreprocessing,suchasscaling.Imageenhancementisamongthesimplestandmostappealingareasofdigitalimageprocessing.Basically,theideabehindenhancementtechniquesistobringoutdetailthatisobscured,orsimplytohighlightcertainfeaturesofinterestinanimage.Afamiliarexampleofenhancementiswhenweincreasethecontrastofanimagebecause“itlooksbetter.”Itisimportanttokeepinmindthatenhancementisaverysubjectiveareaofimageprocessing.Imagerestorationisanareathatalsodealswithimprovingtheappearanceofanimage.However,unlikeenhancement,whichissubjective,imagerestorationisobjective,inthesensethatrestorationtechniquestendtobebasedonmathematicalorprobabilisticmodelsofimagedegradation.Enhancement,ontheotherhand,isbasedonhumansubjectivepreferencesregardingwhatconstitutesa“good”enhancementresult.ColorimageprocessingisanareathathasbeengaininginimportancebecauseofthesignificantincreaseintheuseofdigitalimagesovertheInternet.Itcoversanumberoffundamentalconceptsincolormodelsandbasiccolorprocessinginadigitaldomain.Colorisusedalsoinlaterchaptersasthebasisforextractingfeaturesofinterestinanimage.Waveletsarethefoundationforrepresentingimagesinvariousdegreesofresolution.Inparticular,thismaterialisusedinthisbookforimagedatacompressionandforpyramidalrepresentation,inwhichimagesaresubdividedsuccessivelyintosmallerregions.Compression,asthenameimplies,dealswithtechniquesforreducingthestoragerequiredtosaveanimage,orthebandwidthrequiredtotransmiit.Althoughstoragetechnologyhasimprovedsignificantlyoverthepastdecade,thesamecannotbesaidfortransmissioncapacity.ThisistrueparticularlyinusesoftheInternet,whicharecharacterizedbysignificantpictorialcontent.Imagecompressionisfamiliar(perhapsinadvertently)tomostusersofcomputersintheformofimagefileextensions,suchasthejpgfileextensionusedintheJPEG(JointPhotographicExpertsGroup)imagecompressionstandard.Morphologicalprocessingdealswithtoolsforextractingimagecomponentsthatareusefulintherepresentationanddescriptionofshape.Thematerialinthischapterbeginsatransitionfromprocessesthatoutputimagestoprocessesthatoutputimageattributes.Segmentationprocedurespartitionanimageintoitsconstituentpartsorobjects.Ingeneral,autonomoussegmentationisoneofthemostdifficulttasksindigitalimageprocessing.Aruggedsegmentationprocedurebringstheprocessalongwaytowardsuccessfulsolutionofimagingproblemsthatrequireobjectstobeidentifiedindividually.Ontheotherhand,weakorerraticsegmentationalgorithmsalmostalwaysguaranteeeventualfailure.Ingeneral,themoreaccuratethesegmentation,themorelikelyrecognitionistosucceed.Representationanddescriptionalmostalwaysfollowtheoutputofasegmentationstage,whichusuallyisrawpixeldata,constitutingeitherthebound-aryofaregion(i.e.,thesetofpixelsseparatingoneimageregionfromanother)orallthepointsintheregionitself.Ineithercase,convertingthedatatoaformsuitableforcomputerprocessingisnecessary.Thefirstdecisionthatmustbemadeiswhetherthedatashouldberepresentedasaboundaryorasacompleteregion.Boundaryrepresentationisappropriatewhenthefocusisonexternalshapecharacteristics,suchascornersandinflections.Regionalrepresentationisappropriatewhenthefocusisoninternalproperties,suchastextureorskeletalshape.Insomeapplications,theserepresentationscomplementeachother.Choosingarepresentationisonlypartofthesolutionfortrans-formingrawdataintoaformsuitableforsubsequentcomputerprocessing.Amethodmustalsobespecifiedfordescribingthedatasothatfeaturesofinterestarehighlighted.Description,alsocalledfeatureselection,dealswithextractingattributesthatresultinsomequantitativeinformationofinterestorarebasicfordifferentiatingoneclassofobjectsfromanother.Recognitionistheprocessthatassignsalabel(e.g.,“vehicle”)toanobjectbasedonitsdescriptors.Asdetailedbefore,weconcludeourcoverageofdigitalimageprocessingwiththedevelopmentofmethodsforrecognitionofindividualobjects.SofarwehavesaidnothingabouttheneedforpriorknowledgeorabouttheinteractionbetweentheknowledgebaseandtheprocessingmodulesinFig2above.Knowledgeaboutaproblemdomainiscodedintoanimageprocessingsystemintheformofaknowledgedatabase.Thisknowledgemaybeassim-pleasdetailingregionsofanimagewheretheinformationofinterestisknowntobelocated,thuslimitingthesearchthathastobeconductedinseekingthatinformation.Theknowledgebasealsocanbequitecomplex,suchasaninterrelatedlistofallmajorpossibledefectsinamaterialsinspectionproblemoranimagedatabasecontaininghigh-resolutionsatelliteimagesofaregionincon-nectionwithchange-detectionapplications.Inadditiontoguidingtheoperationofeachprocessingmodule,theknowledgebasealsocontrolstheinteractionbetweenmodules.ThisdistinctionismadeinFig2abovebytheuseofdouble-headedarrowsbetweentheprocessingmodulesandtheknowledgebase,asop-posedtosingle-headedarrowslinkingtheprocessingmodules.EdgedetectionEdgedetectionisaterminologyinimageprocessingandcomputervision,particularlyintheareasoffeaturedetectionandfeatureextraction,torefertoalgorithmswhichaimatidentifyingpointsinadigitalimageatwhichtheimagebrightnesschangessharplyormoreformallyhasdiscontinuities.Althoughpointandlinedetectioncertainlyareimportantinanydiscussiononsegmentation,edgedectectionisbyfarthemostcommonapproachfordetectingmeaningfuldiscountiesingraylevel.Althoughcertainliteraturehasconsideredthedetectionofidealstepedges,theedgesobtainedfromnaturalimagesareusuallynotatallidealstepedges.Insteadtheyarenormallyaffectedbyoneorseveralofthefollowingeffects:1.focalblurcausedbyafinitedepth-of-fieldandfinitepointspreadfunction;2.penumbralblurcausedbyshadowscreatedbylightsourcesofnon-zeroradius;3.shadingatasmoothobjectedge;4.localspecularitiesorinterreflectionsinthevicinityofobjectedges.Atypicaledgemightforinstancebetheborderbetweenablockofredcolorandablockofyellow.Incontrastaline(ascanbeextractedbyaridgedetector)canbeasmallnumberofpixelsofadifferentcoloronanotherwiseunchangingbackground.Foraline,theremaythereforeusuallybeoneedgeoneachsideoftheline.Toillustratewhyedgedetectionisnotatrivialtask,letusconsidertheproblemofdetectingedgesinthefollowingone-dimensionalsignal.Here,wemayintuitivelysaythatthereshouldbeanedgebetweenthe4thand5thpixels.

5

7

6

4

152

148

149

Iftheintensitydifferenceweresmallerbetweenthe4thandthe5thpixelsandiftheintensitydifferencesbetweentheadjacentneighbouringpixelswerehigher,itwouldnotbeaseasytosaythatthereshouldbeanedgeinthecorrespondingregion.Moreover,onecouldarguethatthiscaseisoneinwhichthereareseveraledges.Hence,tofirmlystateaspecificthresholdonhowlargetheintensitychangebetweentwoneighbouringpixelsmustbeforustosaythatthereshouldbeanedgebetweenthesepixelsisnotalwaysasimpleproblem.Indeed,thisisoneofthereasonswhyedgedetectionmaybeanon-trivialproblemunlesstheobjectsinthesceneareparticularlysimpleandtheilluminationconditionscanbewellcontrolled.Therearemanymethodsforedgedetection,butmostofthemcanbegroupedintotwocategories,search-basedandzero-crossingbased.Thesearch-basedmethodsdetectedgesbyfirstcomputingameasureofedgestrength,usuallyafirst-orderderivativeexpressionsuchasthegradientmagnitude,andthensearchingforlocaldirectionalmaximaofthegradientmagnitudeusingacomputedestimateofthelocalorientationoftheedge,usuallythegradientdirection.Thezero-crossingbasedmethodssearchforzerocrossingsinasecond-orderderivativeexpressioncomputedfromtheimageinordertofindedges,usuallythezero-crossingsoftheLaplacianorthezero-crossingsofanon-lineardifferentialexpression,aswillbedescribedinthesectionon\o"Edgedetection"differentialedgedetectionfollowingbelow.Asapre-processingsteptoedgedetection,asmoothingstage,typicallyGaussiansmoothing,isalmostalwaysapplied(seealsonoisereduction).Theedgedetectionmethodsthathavebeenpublishedmainlydifferinthetypesofsmoothingfiltersthatareappliedandthewaythemeasuresofedgestrengtharecomputed.Asmanyedgedetectionmethodsrelyonthecomputationofimagegradients,theyalsodifferinthetypesoffiltersusedforcomputinggradientestimatesinthex-andy-directions.Oncewehavecomputedameasureofedgestrength(typicallythegradientmagnitude),thenextstageistoapplyathreshold,todecidewhetheredgesarepresentornotatanimagepoint.Thelowerthethreshold,themoreedgeswillbedetected,andtheresultwillbeincreasinglysusceptibletonoise,andalsotopickingoutirrelevantfeaturesfromtheimage.Converselyahighthresholdmaymisssubtleedges,orresultinfragmentededges.Iftheedgethresholdingisappliedtojustthegradientmagnitudeimage,theresultingedgeswillingeneralbethickandsometypeofedgethinningpost-processingisnecessary.Foredgesdetectedwithnon-maximumsuppressionhowever,theedgecurvesarethinbydefinitionandtheedgepixelscanbelinkedintoedgepolygonbyanedgelinking(edgetracking)procedure.Onadiscretegrid,thenon-maximumsuppressionstagecanbeimplementedbyestimatingthegradientdirectionusingfirst-orderderivatives,thenroundingoffthegradientdirectiontomultiplesof45degrees,andfinallycomparingthevaluesofthegradientmagnitudeintheestimatedgradientdirection.Acommonlyusedapproachtohandletheproblemofappropriatethresholdsforthresholdingisbyusingthresholdingwithhysteresis.Thismethodusesmultiplethresholdstofindedges.Webeginbyusingtheupperthresholdtofindthestartofanedge.Oncewehaveastartpoint,wethentracethepathoftheedgethroughtheimagepixelbypixel,markinganedgewheneverweareabovethelowerthreshold.Westopmarkingouredgeonlywhenthevaluefallsbelowourlowerthreshold.Thisapproachmakestheassumptionthatedgesarelikelytobeincontinuouscurves,andallowsustofollowafaintsectionofanedgewehavepreviouslyseen,withoutmeaningthateverynoisypixelintheimageismarkeddownasanedge.Still,however,wehavetheproblemofchoosingappropriatethresholdingparameters,andsuitablethresholdingvaluesmayvaryovertheimage.Someedge-detectionoperatorsareinsteadbaseduponsecond-orderderivativesoftheintensity.Thisessentiallycapturestherateofchangeintheintensitygradient.Thus,intheidealcontinuouscase,detectionofzero-crossingsinthesecondderivativecaptureslocalmaximainthegradient.Wecancometoaconclusionthat,tobeclassifiedasameaningfuledgepoint,thetransitioningraylevelassociatedwiththatpointhastobesignificantlystrongerthanthebackgroundatthatpoint.Sincewearedealingwithlocalcomputations,themethodofchoicetodeterminewhetheravalueis“significant”ornotidtouseathreshold.Thuswedefineapointinanimageasbeingasbeinganedgepointifitstwo-dimensionalfirst-orderderivativeisgreaterthanaspecifiedcriterionofconnectednessisbydefinitionanedge.Thetermedgesegmentgenerallyisusediftheedgeisshortinrelationtothedimensionsoftheimage.Akeyprobleminsegmentationistoassembleedgesegmentsintolongeredges.Analternatedefinitionifweelecttousethesecond-derivativeissimplytodefinetheedgeponitsinanimageasthezerocrossingsofitssecondderivative.Thedefinitionofanedgeinthiscaseisthesameasabove.Itisimportanttonotethatthesedefinitionsdonotguaranteesuccessinfindingedgeinanimage.Theysimplygiveusaformalismtolookforthem.First-orderderivativesinanimagearecomputedusingthegradient.Second-orderderivativesareobtainedusingtheLaplacian.数字图像处理与边缘检测数字图像处理数字图像处理方法的研究源于两个主要应用领域：其一是为了便于人们分析而对图像信息进行改良：其二是为使机器自动理解而对图像数据进行存储、传输及显示。一幅图像可定义为一个二维函数f(x,y)，这里x和y是空间坐标，而在任何一对空间坐标〔x,y〕上的幅值f称为该点图像的强度或灰度。当x,y和幅值f为有限的、离散的数值时，称该图像为数字图像。数字图像处理是指借用数字计算机处理数字图像，值得提及的是数字图像是由有限的元素组成的，每一个元素都有一个特定的位置和幅值，这些元素称为图像元素、画面元素或像素。像素是广泛用于表示数字图像元素的词汇。视觉是人类最高级的感知器官，所以，毫无疑问图像在人类感知中扮演着最重要的角色。然而，人类感知只限于电磁波谱的视觉波段，成像机器那么可覆盖几乎全部电磁波谱，从伽马射线到无线电波。它们可以对非人类习惯的那些图像源进行加工，这些图像源包括超声波、电子显微镜及计算机产生的图像。因此，数字图像处理涉及各种各样的应用领域。图像处理涉及的范畴或其他相关领域〔例如，图像分析和计算机视觉〕的界定在初创人之间并没有一致的看法。有时用处理的输入和输出内容都是图像这一特点来界定图像处理的范围。我们认为这一定义仅是人为界定和限制。例如，在这个定义下，甚至最普通的计算一幅图像灰度平均值的工作都不能算做是图像处理。另一方面，有些领域〔如计算机视觉〕研究的最高目标是用计算机去模拟人类视觉，包括理解和推理并根据视觉输入采取行动等。这一领域本身是人工智能的分支，其目的是模仿人类智能。人工智能领域处在其开展过程中的初期阶段，它的开展比预期的要慢的多，图像分析〔也称为图像理解〕领域那么处在图像处理和计算机视觉两个学科之间。从图像处理到计算机视觉这个连续的统一体内并没有明确的界线。然而，在这个连续的统一体中可以考虑三种典型的计算处理〔即低级、中级和高级处理〕来区分其中的各个学科。低级处理涉及初级操作，如降低噪声的图像预处理，比照度增强和图像锋利化。低级处理是以输入、输出都是图像为特点的处理。中级处理涉及分割〔把图像分为不同区域或目标物〕以及缩减对目标物的描述，以使其更适合计算机处理及对不同目标的分类〔识别〕。中级图像处理是以输入为图像，但输出是从这些图像中提取的特征〔如边缘、轮廓及不同物体的标识等〕为特点的。最后，高级处理涉及在图像分析中被识别物体的总体理解，以及执行与视觉相关的识别函数〔处在连续统一体边缘〕等。根据上述讨论，我们看到，图像处理和图像分析两个领域符合逻辑的重叠区域是图像中特定区域或物体的识别这一领域。这样，在研究中，我们界定数字图像处理包括输入和输出均是图像的处理，同时也包括从图像中提取特征及识别特定物体的处理。举一个简单的文本自动分析方面的例子来具体说明这一概念。在自动分析文本时首先获取一幅包含文本的图像，对该图像进行预处理，提取〔分割〕字符，然后以适合计算机处理的形式描述这些字符，最后识别这些字符，而所有这些操作都在本文界定的数字图像处理的范围内。理解一页的内容可能要根据理解的复杂度从图像分析或计算机视觉领域考虑问题。这样，我们定义的数字图像处理的概念将在有特殊社会和经济价值的领域内通用。数字图像处理的应用领域多种多样，所以文本在内容组织上尽量到达该技术应用领域的广度。阐述数字图像处理应用范围最简单的一种方法是根据信息源来分类〔如可见光、X射线，等等〕。在今天的应用中，最主要的图像源是电磁能谱，其他主要的能源包括声波、超声波和电子〔以用于电子显微镜方法的电子束形式〕。建模和可视化应用中的合成图像由计算机产生。建立在电磁波谱辐射根底上的图像是最熟悉的，特别是X射线和可见光谱图像。电磁波可定义为以各种波长传播的正弦波，或者认为是一种粒子流，每个粒子包含一定〔一束〕能量，每束能量成为一个光子。如果光谱波段根据光谱能量进行分组，我们会得到下列图1所示的伽马射线〔最高能量〕到无线电波〔最低能量〕的光谱。如下图的加底纹的条带表达了这样一个事实，即电磁波谱的各波段间并没有明确的界线，而是由一个波段平滑地过渡到另一个波段。图像获取是第一步处理。注意到获取与给出一幅数字形式的图像一样简单。通常，图像获取包括如设置比例尺等预处理。图像增强是数字图像处理最简单和最有吸引力的领域。根本上，增强技术后面的思路是显现那些被模糊了的细节，或简单地突出一幅图像中感兴趣的特征。一个图像增强的例子是增强图像的比照度，使其看起来好一些。应记住，增强是图像处理中非常主观的领域，这一点很重要。图像复原也是改良图像外貌的一个处理领域。然而，不像增强，图像增强是主观的，而图像复原是客观的。在某种意义上说，复原技术倾向于以图像退化的数学或概率模型为根底。另一方面，增强以怎样构成好的增强效果这种人的主观偏爱为根底。彩色图像处理已经成为一个重要领域，因为基于互联网的图像处理应用在不断增长。就使得在彩色模型、数字域的彩色处理方面涵盖了大量根本概念。在后续开展，彩色还是图像中感兴趣特征被提取的根底。小波是在各种分辨率下描述图像的根底。特别是在应用中，这些理论被用于图像数据压缩及金字塔描述方法。在这里，图像被成功地细分为较小的区域。压缩，正如其名称所指的意思，所涉及的技术是减少图像的存储量，或者在传输图像时降低频带。虽然存储技术在过去的十年内有了很大改良，但对传输能力我们还不能这样说，尤其在互联网上更是如此，互联网是以大量的图片内容为特征的。图像压缩技术对应的图像文件扩展名对大多数计算机用户是很熟悉的〔也许没注意〕，如JPG文件扩展名用于JPEG〔联合图片专家组〕图像压缩标准。形态学处理设计提取图像元素的工具，它在表现和描述形状方面非常有用。这一章的材料将从输出图像处理到输出图像特征处理的转换开始。分割过程将一幅图像划分为组成局部或目标物。通常，自主分割是数字图像处理中最为困难的任务之一。复杂的分割过程导致成功解决要求物体被分别识别出来的成像问题需要大量处理工作。另一方面，不健壮且不稳定的分割算法几乎总是会导致最终失败。通常，分割越准确，识别越成功。表示和描述几乎总是跟随在分割步骤的输后边，通常这一输出是未加工的数据，其构成不是区域的边缘〔区分一个图像区域和另一个区域的像素集〕就是其区域本身的所有点。无论哪种情况，把数据转换成适合计算机处理的形式都是必要的。首先，必须确定数据是应该被表现为边界还是整个区域。当注意的焦点是外部形状特性〔如拐角和曲线〕时，那么边界表示是适宜