畢業(yè)設(shè)計的翻譯英文文獻關(guān)于圖像分割pdf范本

年4月19日畢業(yè)設(shè)計的翻譯英文文獻關(guān)于圖像分割pdf文檔僅供參考MicroprocessorsandMicrosystems36()215–231ContentslistsavailableatSciVerseScienceDirectMicroprocessorsandMicrosystemsjournalhomepage:/locate/micproAnembeddedsoftware-reconfigurablecolorsegmentationarchitectureforimageprocessingsystemsGrigoriosChrysosa,?,ApostolosDollasa,NikolaosBourbakisa,baTechnicalUniversityofCrete,ECEDept.,Chania,Crete,GreecebWrightStateUniversity,Engr.CollegeATRCenter,Dayton,OH45435,USAarticle infoArticlehistory:Availableonline17DecemberKeywords:ReconfigurablearchitecturesImagesegmentationEmbeddedsystems

abstractImagesegmentationisoneofthefirstimportantanddifficultstepsofimageanalysisandcomputervisionanditisconsideredasoneoftheoldestproblemsinmachinevision.Lately,severalsegmentationalgorithmshavebeendevelopedwithfeaturesrelatedtothresholding,edgelocationandregiongrowingtoofferanopportunityforthedevelopmentoffasterimage/videoanalysisandrecognitionsystems.Inaddition,fuzzy-basedsegmentationalgorithmshaveessentiallycontributedtosynthesisofregionsforbet-terrepresentationofobjects.Thesealgorithmshaveminordifferencesintheirperformanceandtheyallperformwell.Thus,theselectionofonealgorithmvs.anotherwillbebasedonsubjectivecriteria,or,drivenbytheapplicationitself.Here,alow-costembeddedreconfigurablearchitecturefortheFuzzy-likereason-ingsegmentation(FRS)methodispresented.TheFRSmethodhasthreestages(smoothing,edgedetectionandtheactualsegmentation).Theinitialsmoothingoperationisintendedtoremovenoise.Thesmootherandedgedetectoralgorithmsarealsoincludedinthisprocessingstep.Thesegmentationalgorithmusesedgeinformationandthesmoothedimagetofindsegmentspresentwithintheimage.InthisworktheFRSsegmentationalgorithmwasselectedduetoitsprovengoodperformanceonavarietyofapplications(facedetection,motiondetection,AutomaticTargetRecognition(ATR))andhasbeendevelopedinalow-cost,reconfigurablecomputingplatform,aimingatlowcostapplications.Inparticular,thispaperpresentstheimplementationofthesmoothing,edgedetectionandcolorsegmentationalgorithmsusingStretchS5000processorsandcomparesthemwithasoftwareimplementationusingtheMatlab.Thenewarchitec-tureispresentedindetailinthiswork,togetherwithresultsfromstandardbenchmarksandcomparisonstoalternativetechnologies.Thisisthefirstsuchimplementationthatweknowof,havingatthesametimehighthroughput,excellentperformance(atleastinstandardbenchmarks)andlowcost.ElsevierB.V.Allrightsreserved.1.Introduction1.1.SegmentationManycomputervision,patternrecognition,imageanalysisandobjectextractionsystemshavebeendevelopedduringthelastthirtyyears.Atthesametime,fuzzyandsemi-fuzzyclusteringalgorithmshavebeenalsopresentedfortheextractionandrecog-nitionofanobject’sfeatures.Inorderforthesesystemsandalgo-rithmstobesuccessfultheygenerallyhavetostartwitharobustsmoothingand/orsegmentationtechnique.Thus,imagesegmenta-tionisanimportantstartingstepforalmostallvisionandpatternrecognitionmethodologies.Severalstudieshavebeendonetocat-egorizesegmentationintoclassesbasedoncharacteristics,suchasthresholdingorclustering,edgedetection,regiongrowing/merging?Correspondingauthor.E-mailaddresses:(G.Chrysos),(A.Dollas),(N.Bourbakis).0141-9331/$-seefrontmatterElsevierB.V.Allrightsreserved.doi:10.1016/j.micpro..12.004

andothers[1–3].Inparticular,LeeandChung[4]showedthatthresholdingwouldusuallyproducegoodresultsinbimodalimagesonly,wheretheimagescompriseofonlyoneobjectanditsbackground.However,whentheobjectareaissmallcomparedtothebackgroundarea,orwhenboththeobjectandbackgroundhaveabroadrangeofgraylevels,selectingagoodthresholdisdif-ficult.Anotherweaknessofthistechniqueoccurswhenmultipleobjectsarepresentwithintheimage.Insuchcases,findingsharpvalleyswithinthehistogramisfurthercomplicated,andsegmenta-tionresultsmaybeverypoor.Edgedetectionisanotherapproachassociatedtoimagesegmentation[5].Anedgeisdefinedasaloca-tionwhereasharpchangeingraylevelorcolorisdetected.How-ever,inthismethoditisdifficulttomaintainthecontinuityofthedetectededges;asegmentmustalwaysbeenclosedbyacontinu-ousedge.Regiongrowingormergingisathirdapproachforimagesegmentation[6].Inthiscase,large,easytofindcontinuousre-gionsorsegmentsaredetectedfirst.Afterwards,smallregionsmaybemergedbyusinghomogeneitycriteria[7,8].Onedisadvan-tageofregiongrowingandmergingistheinherentlysequential216 G.Chrysosetal./MicroprocessorsandMicrosystems36()215–231natureofthisapproach.Often,theregionsproduceddependupontheorderinwhichthoseregionsgrowormerge.1.2.ColorsegmentationarchitecturesTheliteraturereportsdifferentapproachesforcolorsegmenta-tion.Animportantcolorsegmentationmethodisthedevelopmentofdichromaticreflectionmodel[15,16],whichdescribesthecolorofreflectedlightasalinearcombinationofthecolorofsurfacereflection(highlights)andbodyreflection(objectcolor).Useofthismodeltotheregiongrowingandmergingmethod[6,17]producedimpressiveresults.Inthismethod,highlightedareasweremergedwiththematteareasofanobject.However,usinghardthresholdsthroughoutdegradedtheperformanceofthistechniquewithinitsintermediatestages.Therearesegmentationmethods[18,19]whichdonotsegmentthecolorimageintheRGBcolorspace,asitdoesnotcloselymodelthepsychologicalunderstandingofcolor.Insteadof,theychooseothercolorspaces,likeHISorYUV,whichproducebetterresultsthantheRGBcolorspace.Someoftheseimagesegmentationpro-cesseswerefusedwiththeedgelocationmethodtoproducebetterresults[20,21].Segmentationbasedonthetheoryofapproximatereasoningorfuzzy-likereasoningproducedpromisingresults[22,23].Huntsberger[5]definedcoloredgesasthezerocrossingofdifferencesbetweenthemembershipvaluesofeachpixel.Thefuzzymembershipvaluesaregeneratedbyusinganiterativec-meansegmentationalgorithmalthoughitistimeconsumingduetoitsiterativenature.Lim[24]presentedanautomatedcoarse-to-finesegmentationmethod.Thisapproachisbasedonhistogramthresholdsforeachcolorandthec-meansalgorithm[25,26].Aninterestingapproach,proposedbyLambertandCarron[27],com-binedthecolorspace(wherehuewasexplicitlydefinedandpro-cessedaccordingtoitsrelevancetochroma)andsymbolicrepresentationsandrule-basedsystems(usingcolorandlumi-nancefeaturestodeterminehomogeneityamongpixels).Recently,moresegmentationtechniquesbasedoncolorandtexturehavebeenintroducedusingfeaturescommonlyobservedinmostimages,especiallyincolortexturedimagesofnaturalscenes.Extensiveresearchresultsonhumanperceptionofcolorandtexturearealsoavailableintheliterature,e.g.,uniformcolorspaces[64]orfilterbanks[35–37].Forallthesereasons,mostseg-mentationmethodsusecolorortextureaskeyfeaturesforimagesegmentation.Recently,severalattemptstocombinecolorandtexturehavebeenmadetoenhancethebasicperformanceofcolorortexturesegmentation.Theseattempts,namelycolor-texturesegmentation,includeregiongrowingapproaches[38–40],wa-tershedtechniques[41],edgeflowtechniques[42],andstochasticmodel-basedapproaches[43,44].TheapplicationofMarkovmod-elsoncolorsegmentationhasalsobeenstudied[45,46].LastlytheBoyokovet.al.[47–49]approachtocolor-texturesegmentationisbasedongraphcuttechniqueswhichfindanoptimalcolor-texturesegmentationofacolortexturedimagebyregardingitasamini-mumcutprobleminaweightedgraph.Therearemanyarchitecturesandhardwareimplementationsofcolorsegmentationalgorithmsinliterature.PerezandKoch[28]proposedtheuseofasimplifiedhuedescriptionsuitableforimple-mentationinanalogVLSI.TheydesignedandfabricatedforthefirsttimeananalogCMOSVLSIcircuitthatcomputesnormalizedcolorandhue.StichlingandKleinjohann[29]presentahardwareimple-mentationofcolorsegmentationalgorithmusingregiongrowingandmergingmethodsimplementedonaPhilipsTrimediamicro-controller.Thesystemprocesses25framespersecrateforsmallimagesandusingaHW-SWsystem.LeclerqandBraunl[30]imple-mentedacolorsegmentationalgorithmona32-bitMotorolacon-trollerfor8060images.ThesystemwasusedfortheRobocupcompetitionandidentifiessmallobjectsinabout0.02s.Saffiotti

[31]presentstheimplementationofaseededregiongrowingseg-mentationalgorithmonaSonyAIBOrobotusingthespecificdeviceCDTthatusesthresholdtechnique.Johnstonetal.[32]presentasystemthatimplementscolorsegmentationandobjecttrackingusinganFPGA(SpartanII)andofferingrealtimeprocessing.Kooetal.[33]presentasystemthatanalyzesmagneticresonanceimages.Thesystemwasimplementedonahigh-performancereconfigurablecomputerusing4FPGAsandachievesa5speedupofthealgorithm.Dillingeretal.[34]builtanFPGA-basedcoproces-sorwhichimplementsa3-Dimagesegmentationachievinghighperformance.Yamaokaetal.[35]presentanovelalgorithmimple-mentedonanFPGAtrackingupto220objectson8060videopictures.1.3.SegmentationforimageprocessingbasedsystemsImageprocessingsystemssuchasAutomaticTargetRecogni-tion(ATR),FaceRecognition,andMotionDetection[14,50–54,62]requirearobustandfastsegmentationalgorithm.Thus,thesesystemsuseaprocessforobjectoffeaturesextractionandrecogni-tionappliedtostillimagesand/orvideo[9–13].Forinstance,anATRsystemconsistofacombinationofalgorithms,suchassmoothing,heuristicsegmentation,edgedetection,thinning,regiongrowing,fractals,etc.,appropriatelyselectedtorecognizetargetsundervariousconditions.Thesealgorithms,especiallythesmoothing,segmentationandedgedetectionconsumeasignifi-cantamountofcomputingtimeneededforthesoftwarecomple-tioninanATRsystem.Colorsegmentationisamuch-studiedproblem[45,57,58],asitisusedinapplicationssuchasfacerecog-nition[55,56].Thus,thecontributionofthisworkisanarchitectureandde-tailedhardwaredesignfortheimplementationofthethreetimeconsumingpartsoftheFRSmethodology(smoothing,edgedetec-tionandcolorsegmentation)[7,8,22,23,36],whichweredevelopedasindependentinhardwareas‘‘blackboxes’’toperformaspecificprocedure.Thefinalresultisanimagedividedintoitsobjectswhicharecoloredwiththesamecolor.ThispieceofinformationcanbeusedbythesubsequentstepsoftheATRsystemtoperformfeatureextractionoftheobjectsintheimage.Thecompletesystemwasfullydesignedinareconfigurableprocessorusingthetechnol-ogyofStretch,Inc.Thisisalow-costtechnologywhichleadstoaneasilyembeddablesubsystem.Aswillbeshowninthispaper,thetightcouplingofanembeddedprocessorwithreconfigurablefab-ricallowsforanefficientimplementationofthealgorithm,how-ever,thevastamountsofdatathatneedtobetransferredbetweenthememory,theprocessor,andthereconfigurablepartposechallengeswhichwillbepresentedin-depthinthiswork.TheStretchcompany[51]hasdevelopedtheseriesofS5000andS6000softwareconfigurableprocessors,whichisbasedontheTensilicacoreRISCprocessorwithasmallembeddedreconfigura-blepart.ThedesignflowcomprisesofsystemdevelopmentinC/C++,profilingofthecode,andmappingitscriticalsectionstothereconfigurablefabricasspecial,hardware-implementedinstructions.TheC/C++languageisusedtoprogramtheS5000pro-cessors.StretchCisaC-likelanguagewhichincludessomeexten-sionsforhardwareimplementation.StretchCistheprogramminglanguagewhichmapsthecriticalpartsofthedesignintherecon-figurablepartsoftheprocessor.Therestofthispaperisorganizedasfollows:Section2de-scribestheFSRsegmentationmethodologythatwasimplemented.Section3describesthenewarchitecture,itsmajorsubsystems,theirinterconnection,anditsmappingontheStretchtechnology.Section4hasperformanceresultsandadetailedcomparisontopreviouslypublishedimplementations.Finally,Section5hassomeconclusionsfromthiswork.G.Chrysosetal./MicroprocessorsandMicrosystems36()215–2312172.TheFRSsegmentationmethodologySegmentationisaprocessusedtofacilitatetheextractionofobjectsthatformanimage.TheFRSmethodology,whichisstudiedinthispaper,consistsofthreesteps(priortotherecognitionit-self):smoothing,edgedetectionandcolorsegmentation.Thedata-flowofsegmentationprocessisdescribedinFig.1.Inthiswork,aswillbeshownbelow,theHIS(hue,intensity,saturation)modelisused,fromoriginalRGBimages,anapproachwhichisquitetypicalandhasbeenshowninliterature(seeSection1)toworkwell.2.1.SmoothingalgorithmTheimagescontainnoiseintroducedeitherbythecameraorbecauseoftheimage’stransmissionoveranoisymedium.Ineithercase,thenoisemustberemovedbeforeanyfurtherimageprocess-ingisapplied.Themostcommonwayofnoiseremovalistheuseoffilters.Animportantconceptforasmoothingalgorithmistheneighborhoodbetweentwopixels.Thisalgorithmallowsforafuz-zydegreeofneighborhood,inwhichforeachneighboringpixelthereisthecorrespondingdegreeofneighborhood,asshowninFig.2.Eachpixel’scoloriscomparedwiththecolorofeachofitsneighboringblocks,asshowninFig.3.Thesizeofblocksforourimplementationwas33,whichresultstoastrongsmoothingoftheimage.TheaveragecolorforeachoftheneighboringblockswascalculatedtakingintoaccounttheneighborhoodmembershipfunctionasshownintheEq.(1).Forsmoothing,thecolorcontrastbetweenthecenterpixelandallofthesurroundingblocksmustbemeasured.Thecolorcontrastbetweenthepixel(i,j)andtheblockbistheEuclideandistanceintheRGBdomainasshowninthefol-lowingequation:pt3kt3q?ps?klsqCsqe1TCi;j;b?PqPpsklsqContrasti;j;b?eR2R1T2teG2G1T2teB2B1T2e2TThestepsofthesmoothingalgorithmthatwereimplementedinthisworkareshowninFig.4andtheyarepresentedanalyticallyin[23].OriginalImage

Fig.2.Tableofneighborhooddegree(lsq).2.2.EdgedetectionEdgedetectionistheprocessofthelimitspecificationoftheob-jectsanimageconsistsof.Hue,IntensityandSaturation(repre-sentedash,i,andsrespectively)areonesetofparametersthatareusedtoevaluatepixels’edgestrengthwithinimages.TheseparametersarecomputedfromtheoriginalimageRGBvaluesbytheequationsbelow:x?0:49rt0:31gt0:2b;y?0:177rt0:812gt0:011b;z?0:01gt0:99byxyl?116;a?500"#;Y0X0Y0"yz#b?200Y0Z0??q????????????????2h?tan1bil;sa2tb;aSmoothedImageEDGESMOOTHINGDETECTIONCOLORSEGMENTATIONColorEdgeImageSegmentedImageFig.1.ThedataflowofFSRalgorithm.218 G.Chrysosetal./MicroprocessorsandMicrosystems36()215–231Fig.3.Eightneighboringblocksofsize3 3andfouredgedirections.Blocksarenumbered1–8suchthattheymaybereferredtoEqs.(1)and(2)(variableb).InitialImageComputecolorcontrastsbetweenthepixelandit’s8neighboringblocksFindmaxandmincolorcontrastMaxcontrast<τsm?NOMincontrast<τsm?

FOREACHPIXELINTHEIMAGEComputeaverageYEScoloraroundthepixelwithablocksizeof7x7ComputeaveragecolorYES betweenthepixelandtheblockwiththeminimumcontrastComputeaveragecolorcontrastbetweenthenorthblockandtheotherthreeblocksplusthecontrastwiththepixelComputeaveragecolorcontrastbetweenthesouthblockandtheotherthreeblocksplusthecontrastwiththepixelComputeaveragecolorcontrastbetweentheeastblockandtheotherthreeblocksplusthecontrastwiththepixelComputeaveragecolorcontrastbetweenthewestblockandtheotherthreeblocksplusthecontrastwiththepixelFindthecolorofthesidewiththelowestcontrastmeasureReplacethepixel’scolorFig.4.Theflowchartofsmoothingalgorithmforpixeli,jandablocksizeof3 3.Inthefirststepsofthealgorithm,thevaluesoftheh,sandiarecomputedforalleight33blocksaroundapixel(Fig.3).Anobjecthasthesamehuethroughout,regardlessofvariancesinshades,high-

lightsandshadows.Ontheotherhand,hueisunstableatlowsatura-tionsandintensities,thereforethehueshouldbenormalized.Thethreevalues(hue,intensityandsaturation)leadtothecalculationG.Chrysosetal./MicroprocessorsandMicrosystems36()215–231219ofthepixelsthatareobjectedgesaccordingtothealgorithmpre-sentedin[23].TheflowdiagramoftheedgedetectionalgorithmisshowninFig.5.2.3.ColorsegmentationalgorithmThesegmentationalgorithmusesedgeinformationandtheinformationofthesmoothedimagetofindsegments.Thestepsin-volvedinthissegmentationprocedurefollow:Findbigandcrispsegments.Expandsegmentsbasedonhomogeneitycriteria.Expandsegmentsbasedonthedichromaticreflectionmodel