翻譯以原文和在同一文件中前

圖像割圖像析計機視第重要最的驟之，也認為機定位和區(qū)域增長，這些為更快的圖像/分析和識別系統(tǒng)的發(fā)展提供了很好的機（FSFS（緣檢測算法也包括在這個程序步驟中。分割算法利用了邊緣信息和平滑后的圖像，用來找到圖像中的區(qū)域。在這項工作中FS分割算法被選擇是依賴于它在一些列應(yīng)用中（（T用StrtchS5000用它的性能。有一些分割方法，它們不在RGB空間里分割彩像，因為它們和HISYUV，這RGB色彩空間顯示出了更好的效果。這些圖像分割過程融合了邊緣檢測方法來產(chǎn)生更好的結(jié)果。基于近似推理或模糊推理的分割產(chǎn)生了可喜的成果。Huntsberger定義顏色的邊緣為每個像素成員函數(shù)值的差異為零。通過使用了C迭代分割算法得到了C迭代算法由于本身性質(zhì)來說是耗時的。Lim自動化的從粗到細的分割方法。這種方法基于閾值直方圖和C迭代算法。Lambert和，緣流技術(shù)，基于模型的隨機分割法。Markov模型的色彩分割應(yīng)用也被研究。最后，Boyokov等人基于圖像分割技術(shù)原理提出了一種顏色-紋理分割方法，把它看做一個加的架構(gòu)和硬件實現(xiàn)。Perez和Koch提出了一種簡化的適合執(zhí)行在模擬集成電的色相描述。他們首次設(shè)計、制造并模擬了CMOS超大規(guī)模集成電路來計算標(biāo)準(zhǔn)化色相和色調(diào)。StichlingKleinjohann提出用硬件實現(xiàn)顏色分割算法，通過使用區(qū)域增長和合并法在飛利浦Trimedia微控制器上實現(xiàn)。這種系統(tǒng)每秒可處理25幀小圖像，使用HW-SW系統(tǒng)。Leclerq和Braun在一個32位的摩托羅拉的控制器上實現(xiàn)了顏色分割80*600.02秒內(nèi)識別備在索尼AIBO機器人身上實現(xiàn)。Johnston等人做了一個系統(tǒng)，使用了FPGA，實現(xiàn)了顏色分割和對象并提供實時處理。Koo等人做出了一個系統(tǒng)，用來分析磁圖4FPGA的計算機上實現(xiàn)，并達到了五倍的加速比。Dillinger等人建立了一個基于FPGA的程序，實現(xiàn)了三維的分的圖像上220個對象。止圖像或。例如，一個ATR系統(tǒng)包括許多算法的組合，啟發(fā)式分割，平緩，邊緣法，特別是平滑、分割和邊緣檢測，在ATR系統(tǒng)中軟件完成所需要的時間里，它們占FRS方法（平滑，邊緣檢這一部分的信息可以被隨后的ATR系統(tǒng)執(zhí)行對象在圖像中的特征提取這一步驟所利以產(chǎn)生一個嵌入式的子系統(tǒng)。如本文所示，會看到一個嵌入式處理器，緊密耦合理器和可重配置的部分之間傳送，將對這項工作的深入提出了。Stretch公司已經(jīng)開發(fā)出S5000和S6000系列軟件配置處理器，這都是基于Tensilica的RISC處C/C++平臺上開發(fā)，包括本2FSR3節(jié)描述新的體5節(jié)會對本工作做出FRS1RGB工作的很好。（見第1節(jié)。圖 FRS算法的數(shù)據(jù)流模糊算子，其中為每個相鄰像素的領(lǐng)域都設(shè)置了相應(yīng)的算子，如圖2所示。圖 333*318，以便帶入下面的兩個式子計我們實施的塊的大小為3*3，這樣對圖像可以達到高度平滑的效果。每一個領(lǐng)域塊的平均顏色都按照下面所示的函數(shù)（1）來進行計算。要做到平滑，必須測量中心像素和所有周圍像素域之間的顏色對比度。像素（i,j）和領(lǐng)域塊b之間的顏色對比，從原始圖像RGB數(shù)值中用這些參數(shù)通過幾何學(xué)來表示為下面的式子：4圖 數(shù)。這些參數(shù)是從原始圖像的RGB值由以下算得的：在該算法的第一個步驟中，h,si3*3（圖3。一個物體，無論是色調(diào)，亮度，或者是陰影的差異，都具有相同的色相貫穿5圖 意義上近鄰近的區(qū)域：近空間域（物理近；或近集群域的顏色立方體（幾乎個過濾器用作三大小分別是3*3，5*5，7*7的塊，最小的一個被最先應(yīng)用。ATR算法的前三個步驟，其中還包括三個子系統(tǒng)，每一個都為一個算法的主要方面來66CStretchS5000StretchC編寫的語言，映射在可重構(gòu)的部分，就是處理器上所謂的ISEF。每個使用可重構(gòu)開發(fā)ISEF上的映射并使用各自的輸入數(shù)據(jù)運行它的代碼。并行代碼完成后的結(jié)果返回到程序，串行代碼繼續(xù)執(zhí)行。實現(xiàn)模塊之間的通信將通過具有內(nèi)部功能的S5000處理器。最初，彩片像素的RGB值在處理器的內(nèi)部中，平滑模塊通過這平滑子系統(tǒng)需要原始像素的RGB值作為輸入和平滑后的圖像數(shù)值作為輸出。該平滑算法分為七個小步驟，每一個都被設(shè)計為一個單獨的模塊。結(jié)構(gòu)如圖7所示。7S5000處理器上并行工83*33所示。這些C(i,j),b7所示。S5000處理器可重構(gòu)部分的限制之一是，用來在可重構(gòu)的部分和處理器的軟件部分之間進行通信的I/O寄存器（128位*3=384位）I/O寄存器的限制每次都會引起相鄰兩個窗口之間的顏色計算，兩個相鄰的3*3窗口包含了15個不同的像素（有一個的3像素，這就意味著15*8*3（像素的顏RGB模式）=360位<384位。這對算子塊對C(i,j),b值的計算如圖3所示。C(i,j),b的計算模塊在可重構(gòu)C(i,j),b的值包含乘法，這是對軟件來C(i,j),b子窗口僅需要三個硬件時在這一階段中，一個新的可重構(gòu)的邏輯模塊被開發(fā)。該模塊把8個鄰域子窗口的C(i,j),b的值，處理的像素的顏色和它們之間的差異計算值作為輸入，完成這個計算后，7所示，第二個是將要被處理的子窗口的ID號。這個過程中每個計算需要9個時鐘間的差異，如圖7所示。計算的過程是一個繁重的問題，因此它發(fā)生在S5000中可像素*3RGB=147*8=1176位)。為了處理這個問題，我們考慮到一個有利條件就是，四C(i,j),b的在第一步的平滑算法中已經(jīng)被可重構(gòu)的硬件計算過了。在這3*37*7的窗口的像素值相加。最后的結(jié)果被7*7窗口的和所相除。C(i,j),b的值和四個窗口的方向（東南西北C(i,j),b的RGB*4個窗口=12*8=96位）作為輸入，這887*7窗口的像素值作為輸入。用顏色平滑單Cbs3*3的子窗口的顏色差異。如圖3所示的四個方向的每一個方向的顏色差異，被暫時起來，被用于色調(diào)，飽和度和強度的每一個方向的計算。最后得出的值被當(dāng)做色調(diào)差異模塊的輸入值在表S5000處理器的可重構(gòu)從原始圖像中產(chǎn)生了。這個信息被在S5000處理器的器中，并且作為下3699顏色分割子系統(tǒng)把S5000處理器器中的邊緣檢測像素值作為輸入，并且把用不同顏色標(biāo)注了對象的，這個可以用于下一步的圖像處理系統(tǒng)，比如 Stretch處理器上的可重構(gòu)部分來執(zhí)行。它的輸入包括處理后的像素值和周圍有3*3窗口的像素值，并且給出信號值顯示該像素是否可以被用作區(qū)集中起來，根據(jù)這個處理的信息把每一個區(qū)域添加一個ID。最后，區(qū)域平均顏色的計Stretch技術(shù)中可能會有限RGB模式上的級數(shù)。另外，軟件乘法器和加法器被設(shè)計用于計算顏色反射值和每一個被擴展后的區(qū)域的顏色。最后的擴展算法FRSStretchFRSSW/HWStretch處理器上的實現(xiàn)。這個算法的S5000FPGA結(jié)構(gòu)上建立，另外這個算法的其他部分被串行的C代碼在嵌入式的Stretch處理器上執(zhí)行。這個小節(jié)描述了這個工作中最重要最的部分，就是Stretch可重構(gòu)處理器能用于實現(xiàn)FSR算法的方如上所述，像素或子窗口程序，和許多算法的不同階段。這些都導(dǎo)致了FPGA上的可度簡化成了定點計算，更快地使可重構(gòu)模塊執(zhí)行。最后，F(xiàn)PGA和處理器之間的數(shù)據(jù)通ContentslistsContentslistsavailableatSciVerseMicroprocessorsandjournal/locate/micproMicroprocessorsandMicrosystems36(2012)Anembeddedsoftware-recon?gurablecolorsegmentationforimageprocessingTechnicalUniversityofCrete,ECEDept.,Chania,Crete,WrightStateUniversity,Engr.CollegeATRCenter,Dayton,OH45435,articleinfArticleAvailableonline17December:Recon?gurablearchitecturesImagesegmentationEmbeddedsystems

abstracImagesegmentationisoneofthe?rstimportantanddif?cultstepsofimage ysisandcomputervisionanditisconsideredasoneoftheoldestproblemsinmachinevision.Lay,severalsegmentationalgorithmshavebeendevelopedwithfeaturesrelatedtothresholding,edgelocationandregiongrowingtoofferanopportunityforthedevelopmentoffasterimage/ ysisandrecognitionsystems.Inaddition,fuzzybasedsegmentationalgorithmshaveessentiallycontributedtosynthesisofregionsforbetterrepresentationofobjects.Thesealgorithmshaveminordifferencesintheirperformanceandtheyallperformwell.Thus,theselectionofonealgorithmvs.anotherwillbebasedonsubjectivecriteria,or,drivenbytheapplicationitself.Here,alowcostembeddedrecon?gurablearchitecturefortheFuzzylikereasoningsegmentation(FRS)methodispresented.TheFRSmethodhasthreestages(smoothing,edgedetectionandtheactualsegmentation).Theinitialsmoothingoperationisintendedtoremovenoise.Thesmootherandedgedetectoralgorithmsarealsoincludedinthisprocessingstep.Thesegmentationalgorithmusesedgeinformationandthesmoothedimageto?ndsegmentspresentwithintheimage.InthisworktheFRSsegmentationalgorithmwasselectedduetoitsprovengoodperformanceonavarietyofapplications(facedetection,motiondetection,AutomaticTargetRecognition(ATR))andhasbeendevelopedinalowcost,recon?gurablecomputingtform,aimingatlowcostapplications.Inparticular,thispaperpresentstheimplementationofthesmoothing,edgedetectionandcolorsegmentationalgorithmsusingStretchS5000processorsandcomparesthemwithasoftwareimplementationusingthe .Thenewarchitectureispresentedindetailinthiswork,togetherwithresultsfromstandardben arksandcomparisonstoalternative.Thisisthe?rstsuchimplementationthatweknowof,havingatthesametimehighthroughput,excellentperformance(atleastinstandardben arks)andlowcost.ó2011ElsevierB.V.All Manycomputervision,patternrecognition,image ysisandobjectextractionsystemshavebeendevelopedduringthelastthirtyyears.Atthesametime,fuzzyandsemifuzzyclusteringalgorithmshavebeenalsopresentedfortheextractionandrecognitionofanobject’sfeatures.Inorderforthesesystemsandalgorithmstobesuccessfultheygenerallyhavetostartwitharobustsmoothingand/orsegmentationtechnique.Thus,imagesegmentationisanimportantstartingstepforalmostallvisionandpatternrecognitionmethodologies.Severalstudieshavebeendonetocategorizesegmentationintoclassesbasedoncharacteristics,suchasthresholdingorclustering,edgedetection,regiongrowing/merging*Correspondingnikolaos.bourbakis@(N.Bourbakis).

andothers[13].Inparticular,LeeandChung[4]showedthatthresholdingwouldusuallyproducegoodresultsinbimodalimagesonly,wheretheimagescompriseofonlyoneobjectanditsbackground.However,whentheobjectareaissmallcomparedtothebackgroundarea,orwhenboththeobjectandbackgroundhaveabroadrangeofgraylevels,selectingagoodthresholdisdif?cult.Anotherweaknessofthistechniqueoccurswhenmultipleobjectsarepresentwithintheimage.Insuchcases,?ndingsharpvalleyswithinthehistogramisfurthercomplicated,andsegmentationresultsmaybeverypoor.Edgedetectionisanotherapproachassociatedtoimagesegmentation[5].Anedgeisde?nedasalocationwhereasharpchangeingraylevelorcolorisdetected.However,inthismethoditisdif?culttomaintainthecontinuityofthedetectededges;asegmentmustalwaysbeenclosedbyacontinuousedge.Regiongrowingormergingisathirdapproachforimagesegmentation[6].Inthiscase,large,easyto?ndcontinuousregionsorsegmentsaredetected?rst.Afterwards,smallregionsmaybemergedbyusinghomogeneitycriteria[7,8].Onedisadvantageofregiongrowingandmergingistheinherentlysequential0141-9331/$-seefrontmatteró2011ElsevierB.V.Allrights G.Chrysosetal./MicroprocessorsandMicrosystems36(2012)natureofthisapproach.Often,theregionsproduceddependupontheorderinwhichthoseregionsgrowormerge.ColorsegmentationTheli turereportsdifferentapproachesforcolorsegmentation.Animportantcolorsegmentationmethodisthedevelopmentofdichromaticre?ectionmodel[15,16],whichdescribesthecolorofre?ectedlightasalinearcombinationofthecolorofsurfacere?ection(highlights)andbodyre?ection(objectcolor).Useofthismodeltotheregiongrowingandmergingmethod[6,17]producedimpressiveresults.Inthismethod,highlightedareasweremergedwiththematteareasofanobject.However,usinghardthresholdsthroughoutdegradedtheperformanceofthistechniquewithinitsintermediatestages.Therearesegmentationmethods[18,19]whichdonotsegmentthecolorimageintheRGBcolorspace,asitdoesnotcloselymodelthepsychologicalunderstandingofcolor.Insteadof,theychooseothercolorspaces,likeHISorYUV,whichproducebetterresultsthantheRGBcolorspace.Someoftheseimagesegmentationprocesseswerefusedwiththeedgelocationmethodtoproducebetterresults[20,21].Segmentationbasedonthetheoryofapproximatereasoningorfuzzylikereasoningproducedpromisingresults[22,23].Huntsberger[5]de?nedcoloredgesasthezerocrossingofdifferencesbetweenthemembershipvaluesofeachpixel.Thefuzzymembershipvaluesaregeneratedbyusingani tivecmeansegmentationalgorithmalthoughitistimeconsumingduetoitsi tivenature.Lim[24]presentedanautomatedcoarseto?nesegmentationmethod.Thisapproachisbasedonhistogramthresholdsforeachcolorandthecmeansalgorithm[25,26].Aninterestingapproach,proposedbyLambertandCarron[27],combinedthecolorspace(wherehuewasexplicitlyde?nedandprocessedaccordingtoitsrelevancetochroma)andsymbolicrepresentationsandrulebasedsystems(usingcolorandluminancefeaturestodeterminehomogeneityamongpixels).Recently,moresegmentationtechniquesbasedoncolorandtexturehavebeenintroducedusingfeaturescommonlyobservedinmostimages,especiallyincolortexturedimagesofnaturalscenes.Extensiveresearchresultsonhumanperceptionofcolorandtexturearealsoavailableintheli ture,e.g.,uniformcolorspaces[64]or?lterbanks[3537].Forallthesereasons,mostsegmentationmethodsusecolorortextureaskeyfeaturesforimagesegmentation.Recently,severalattemptstocombinecolorandtexturehavebeenmadetoenhancethebasicperformanceofcolorortexturesegmentation.Theseattempts,namelycolortexturesegmentation,includeregiongrowingapproaches[3840],watershedtechniques[41],edge?owtechniques[42],andstochasticmodelbasedapproaches[43,44].TheapplicationofMarkovmodelsoncolorsegmentationhasalsobeenstudied[45,46].LastlytheBoyokovet.al.[4749]approachtocolortexturesegmentationisbasedongraphcuttechniqueswhich?ndanoptimalcolortexturesegmentationofacolortexturedimagebyregardingitasaminimumcutprobleminaweightedgraph.Therearemanyarchitecturesandhardwareimplementationsofcolorsegmentationalgorithmsinli ture.PerezandKoch[28]proposedtheuseofasimpli?edhuedescriptionsuitableforimplementationin ogVLSI.Theydesignedandfabricatedforthe?rsttimean ogCMOSVLSIcircuitthatcomputesnormalizedcolorandhue.StichlingandKleinjohann[29]presentahardwareimplementationofcolorsegmentationalgorithmusingregiongrowingandmergingmethodsimplementedonaPhilipsTrimediamicrocontroller.Thesystemprocesses25framespersecrateforsmallimagesandusingaHWSWsystem.LeclerqandBraunl[30]implementedacolorsegmentationalgorithmona32bitMotorolacontrollerfor8060images.ThesystemwasusedfortheRobocupcompetitionandidenti?essmallobjectsinabout0.02s.Saf?otti

[31]presentstheimplementationofaseededregiongrowingsegmentationalgorithmonaSonyAIBOrobotusingthespeci?cdeviceCDTthatusesthresholdtechnique.Johnstonetal.[32]presentasystemthatimplementscolorsegmentationandobjecttrackingusinganFPGA(SpartanII)andofferingrealtimeprocessing.Kooetal.[33]presentasystemthatyzesmagneticresonanceimages.Thesystemwasimplementedonahighperformancerecon?gurablecomputerusing4FPGAsandachievesa5speedupofthealgorithm.Dillingeretal.[34]builtanFPGAbasedcoprocessorwhichimplementsa3Dimagesegmentationachievinghighperformance.Yamaokaetal.[35]presentanovelalgorithmimplementedonanFPGAtrackingupto220objectson8060SegmentationforimageprocessingbasedImageprocessingsystemssuchasAutomaticTargetRecognition(ATR),FaceRecognition,andMotionDetection[14,5054,62]requirearobustandfastsegmentationalgorithm.Thus,thesesystemsuseaprocessforobjectoffeaturesextractionandrecognitionappliedtostillimagesand/or [913].Forinstance,anATRsystemconsistofacombinationofalgorithms,suchassmoothing,heuristicsegmentation,edgedetection,thinning,regiongrowing,fractals,etc.,appropriayselectedtorecognizetargetsundervariousconditions.Thesealgorithms,especiallythesmoothing,segmentationandedgedetectionconsumeasigni?cantamountofcomputingtimeneededforthesoftwarecompletioninanATRsystem.Colorsegmentationisamuchstudiedproblem[45,57,58],asitisusedinapplicationssuchasfacerecognition[55,56].Thus,thecontributionofthisworkisanarchitectureanddetailedhardwaredesignfortheimplementationofthethreetimeconsumingpartsoftheFRSmethodology(smoothing,edgedetectionandcolorsegmentation)[7,8,22,23,36],whichweredevelopedasindependentinhardwareas‘‘blackboxes’’toperformaspeci?cprocedure.The?nalresultisanimagedividedintoitsobjectswhicharecoloredwiththesamecolor.ThispieceofinformationcanbeusedbythesubsequentstepsoftheATRsystemtoperformfeatureextractionoftheobjectsintheimage.Thecompletesystemwasfullydesignedinarecon?gurableprocessorusingthetechnologyofStretch,Inc.Thisisalowcosttechnologywhichleadstoaneasilyembeddablesubsystem.Aswillbeshowninthispaper,thetightcouplingofanembeddedprocessorwithrecon?gurablefabricallowsforanef?cientimplementationofthealgorithm,however,thevastamountsofdatathatneedtobetransferredbetweenthememory,theprocessor,andtherecon?gurablepartposechallengeswhichwillbepresentedindepthinthiswork.The [51]hasdevelopedtheseriesofS5000S6000softwarecon?gurableprocessors,whichisbasedontheTensilicacoreRISCprocessorwithasmallembeddedrecon?gurablepart.Thedesign?owcomprisesofsystemdevelopmentinC/C++,pro?lingofthecode,andmapitscriticalsectionstotherecon?gurablefabricasspecial,hardwareimplementedinstructions.TheC/C++languageisusedtoprogramtheS5000processors.StretchCisaClikelanguagewhichincludessomeextensionsforhardwareimplementation.StretchCistheprogramminglanguagewhichmapsthecriticalpartsofthedesignintherecon?gurablepartsoftheprocessor.Therestofthispaperisorganizedasfollows:Section2describestheFSRsegmentationmethodologythatwasimplemented.Section3describesthenewarchitecture,itsmajorsubsystems,theirinterconnection,anditsmapontheStretchtechnology.Section4hasperformanceresultsandadetailedcomparisontopreviouslypublishedimplementations.Finally,Section5hassomeconclusionsfromthiswork.G.Chrysosetal./MicroprocessorsandMicrosystems36(2012) TheFRSsegmentationSegmentationisaprocessusedtofacilitatetheextractionofobjectsthatformanimage.TheFRSmethodology,whichisstudiedinthispaper,consistsofthreesteps(priortotherecognitionitself):smoothing,edgedetectionandcolorsegmentation.Thedata?owofsegmentationprocessisdescribedinFig.1.Inthiswork,aswillbeshownbelow,theHIS(hue,intensity,saturation)modelisused,fromoriginalRGBimages,anapproachwhichisquitetypicalandhasbeenshowninli ture(seeSection1)toworkwell.SmoothingTheimagescontainnoiseintroducedeitherbythecameraorbecauseoftheimage’stransmissionoveranoisymedium.Ineithercase,thenoisemustberemovedbeforeanyfurtherimageprocessingisapplied.Themostcommonwayofnoiseremovalistheuseof?lters.Animportantconceptforasmoothingalgorithmistheneighborhoodbetweentwopixels.Thisalgorithmallowsforafuzzydegreeofneighborhood,inwhichforeachneighboringpixelthereisthecorrespondingdegreeofneighborhood,asshowninFig.2.Eachpixel’scoloriscomparedwiththecolorofeachofitsneighboringblocks,asshowninFig.3.Thesizeofblocksforourimplementationwas33,whichresultstoastrongsmoothingoftheimage.TheaveragecolorforeachoftheneighboringblockswascalculatedtakingintoaccounttheneighborhoodmembershipfunctionasshownintheEq.(1).Forsmoothing,thecolorcontrastbetweenthecenterpixelandallofthesurroundingblocksmustbemeasured.Thecolorcontrastbetweenthepixel(i,j)andtheblockbistheEuclideandistanceintheRGB asshowninthefol

Fig.2.TableofneighborhooddegreeEdgeEdgedetectionistheprocessofthelimitspeci?cationoftheobjectsanimageconsistsof.Hue,IntensityandSaturation(representedash,i,andsrespectively)areonesetofparametersthatareusedtoevaluatepixels’edgestrengthwithinimages.TheseparametersarecomputedfromtheoriginalimageRGBvaluesbytheequationsbelow: 0:49rt0:31gt0:2b; 0:177rt0:812gt 0:01gtPlowingP

kt3

116

xX X0

y

qPP

kt3

yqpsk

z

RT2t GT2t B

ThestepsofthesmoothingalgorithmthatwereimplementedinthisworkareshowninFig.4andtheyarepresentedyticallyin[23].

l;

a2tb2; tan1aOriginal

EdgeFig.1.Thedata?owofFSR G.Chrysosetal./MicroprocessorsandMicrosystems36(2012)Fig.3.Eightneighboringblocksofsize33andfouredgedirections.Blocksarenumbered1–8suchthattheymaybereferredtoEqs.(1)and(2)(variableComputecolorpixelandit’s8neighboringblocksFOREACHPIXELINTHEIMAGEmincolorMaxcontrast<Mincontrast<FindthecolorofthesidewiththelowestcontrastmeasureComputeaveragecolorcontrastbetweenthenorthblockandtheotherthreeblocksplusthecontrastwiththepixelComputeaveragecolorcontrastbetweenthewestblockandtheotherthreeblocksplusthecontrastwiththepixelComputeaveragecolorcontrastbetweenthesouthblockandtheotherthreeblocksplusthecontrastwiththepixelComputeaveragecolorcontrastbetweentheeastblockandtheotherthreeblocksplusthecontrastwiththepixelcoloraroundthepixelwithablocksizeof7x7ComputeaveragebetweenthepixelandtheblockwiththeminimumcontrastReceFig.4.The?owchartofsmoothingalgorithmforpixeli,jandablocksizeof3Inthe?rststepsofthealgorithm,thevaluesoftheh,sandiarecomputedforalleight33blocksaroundapixel(Fig.3).Anobjecthasthesamehuethroughout,regardlessofvariancesinshades,high

tionsandintensities,thereforethehueshouldbenormalized.Thethreevalues(hue,intensityandsaturation)leadtothecalculationG.Chrysosetal./MicroprocessorsandMicrosystems36(2012) ofthepixelsthatareobjectedgesaccordingtothealgorithmpresentedin[23].The?owdiagramoftheedgedetectionalgorithmisshowninFig.5.ColorsegmentationThesegmentationalgorithmusesedgeinformationandtheinformationofthesmoothedimageto?ndsegments.Thestepsinvolvedinthissegmentationprocedurefollow:FindbigandcrispExpandsegmentsbasedonhomogeneityExpandsegmentsbasedonthedichromaticre?ectionExpandsegmentsbasedonthedegreeoffarnessApplyan tiveThe?rststepofthecolorsegmentationalgorithmistheof?ndingbigandcrispsegments.Onceedgedetectionhas

performedonanimage,crispsegmentsaresurroundedbyedgepixelsortheimageboundary.Crispsegmentscanbede?nedasasetofpixelscompleysurroundedbyedgepixelsbelongingtoonlyoneobject.Thenextstepofthesegmentationalgorithmistheexpansionofthesegmentsbasedonspeci?ccriteriaofhomogeny.Theinitialimageisscannedandusingtheinformationthatresultedfromtheedgedetectiontheexistingsegmentsareexpandedbyaddingpixelswithhighsimilaritytothoseoftheexistingsegments.Thethirdstepofthecolorsegmentationprocessisthesegmentexpansionbasedonthedichromaticre?ectionmodel.Usingthedichromaticre?ectionmodel,someadjacentpixelsmaybemergedwiththepreviouslygrowingmattesegmentsaccordingtoafuzzymeasuresuchasthecustomizeddistancebetweenthemergingpixelandaclusterneinthecolor Tofurtherexpandsegments,the‘‘degreeoffarness’’measureisused.Anunassignedpixelcanbeclose(notfar)toaneighboringsegmentintwosenses:closeinthespatial FindthresholdedumsinthedirectionstoredforeachpixelCalculatethesaturation,intensityandhuecontrastforthe8neighboringForeachoneofCalculatethesaturation,intensityandhuecontrastforthe8neighboringForeachoneof4edgedirectionscalculatetheaveragesaturations,intensitiesandhuecontastsForeachoneof4edgedirectionscalculateμsandμiusingthelowmembershipfunctionCalculatetheaverageofthenormalizedhuecontrastandthefourhuecontrastforeachedgedirectiontl>Maxedgecandidacy>andfindCalculatethefouredgecandidacymeasureshueCalculateEvaluatefourbyoneMergethefouredgestrength G.Chrysosetal./MicroprocessorsandMicrosystems36(2012)close);orcloseinthecluster ofthecolorcube(almostofthesamecolor).Thedegreeoffarnessofapixeltoaneighboringsegmentisde?nedastheproductofthesetwomeasures.Speci?cally,thedegreeoffarnessforanygivenpixelthatwasusedistheabsolutecolorcontrastmultipliedbythegeometricdistance(inpixels)betweenthegivenpixelandthesegmentborder.Finally,afterthesegmentexpansioniscomplete,theresultingsegments’edgesaresmoothedusinganitive?lter.This?lterisusedforthreeblocksizesof33,55and77,withthesmallestonebeingapplied?rst.Inthe?nalimageallthepixelsaregroupedinsegments,accordingtothedescribedcriteria.Eachsegmentoftheimageisconsideredtobeanindependentobjectandiscoloredwiththesamecolor.Thecolorofeachimageiscalculatedastheaveragecolorofthepixelsthatbelongtotheimage.Thealgorithmofcolorsegmentationispresentedindetailin[8].Anarchitectureforsmoothing,edgedetectionandThissectiondescribesthearchitecturesofthesmoothing,edgedetection,andcolorsegmentationalgorithmsthatweredevelopedusingtheStretchtechnology.Thesystem,whichimplementsthe?rstthreestepsoftheATRalgorithm,consistsofthreesubsystems,oneforeachofthemainaspectsofthealgorithm.ThesystemblockdiagramispresentedinFig.6,andthearchitectureofeachsubsystemwillbedescribedinthefollowingsectionsofthispaper.EachsubsystemconsistsofserialCcodewhichrunsontheStretchS5000processorand‘‘hardware’’functions,writteninStretchClanguage,whicharemappedontherecon?gurablepart,thesocalledISEF,oftheprocessor.Eachrecon?gurablefunctionisdevelopedusingtherecon?gurableresourcesoftheprocessoranditscodeisexecutedinparallel.Eachtimetheserialprogramcallsa‘‘hardware’’function,theprocessorloadsitshardwaremapontheISEFandrunsitscodeusingtherespectiveinputdata.Aftertheparallelcodehas?nishedtheresultsreturntotheprocessorandtheserialcodeexecutioncontinues.ThecommunicationbetweentheimplementedcomponentstakescethroughtheinternalmemoryoftheS5000processor.Initially,thevaluesofcolorofthepictures’pixelsinRGBarestoredintheinternalmemoryoftheprocessor.Thesmoothingcomponentreadsthesevalueswhichareprocessedandtheyarestoredagaininthememoryofthesystem.Theprocesscontinuesfortheothertwosystem’scomponentswhichreadtheinputdatafromthememoryandstoretheprocesseddata.Finally,thesystemoutputsthesegmentedimagewherethedetectedobjectsofthepicturearecoloredwiththesamecolorinRGB.SmoothingsubsystemThesmoothingsubsystemtakesasinputtheinitialvaluesofpixelsintheRGBsystemandoutputsthevaluesofthesmoothedimage.Thesmoothingalgorithmisdividedintosevensmallsteps,

eachofwhichwasdesignedasaseparatecomponentasshowninblockdiagraminFig.7.Thedarkcomponentsoftheblockdiagramhavebeendevelopedintherecon?gurablepartofS5000processorandtheycanworkinparallelinordertoreducethetotalexecutiontimeofthesmoothingalgorithm.Initially,accordingtothesmoothingalgorithm,thereare8neighboring33windowsofeachpixel,asshowninFig.3.Thecolorofthesewindows,C(i,j),bvalues,areusedto?ndthecolorcontrastbetweentheprocessingpixelandthesurroundwindows,accordingtoEq.(2).ThecalculationoftheC(i,j),btakesceinrecon?gurablepartoftheprocessor,asshowninFig.7.Oneoftherestrictionsoftherecon?gurablepartofS5000processorsisthesmallnumberofI/Oregisters(128bits3=384bitsatmost)whichareusedforthecommunicationbetweentherecon?gurableandthesoftwarepartoftheprocessor.TheI/Orestrictionledtothecalculationofthecolorfortwoneighboringsubwindowseachtime,as2neighboring33windowscontain15differentpixels(thereisanoverlapof3pixels)whichmeans158bits/pixel3(asthecolorofthepixelsisinRGBmodel)=360bits<384bits.ThepairsofthesubwindowsthatcalculatedthevaluesofC(i,j),b,asshowninFig.3,are53,82,46and17.Also,thedirectionofthesubwindowsisimportantforthecalculationofthewindows’color.Thisledtotheimplementationoftwodifferentcomponentsforthecalculationofthesubwindows,oneforthehorizontalandanotherfortheverticaldirection.ThedesignC(i,j),b’scalculationmoduleonrecon?gurablefabricwasmandatedbytheobservationsthat:(i)ThecalculationoftheC(i,j),bvaluecontainsmultiplications,whichisa‘‘heavy’’taskforthesoftwareand(ii)thisimplementationneedsonlythreehardwareclockcyclesforeachpairoftheC(i,j),b’ssubwindows.Finally,itisimportanttomentionthatastherecon?gurablepartoftheprocessordoesnotsupport?oatingpointarithmetic,?xedpointarithmeticwasusedforthevaluesoftheweightedtable.Thenextstepofthesmoothingalgorithmisthecalculationofthecolorcontrastbetweeneachofneighboringwindowsandtheprocessedpixel,usingtheEq.(3).Inthisstage,anewcomponentofrecon?gurablelogicwasdeveloped.ThiscomponenttakesasinputtheC(i,j),bvaluesofthe8neighboringsubwindowsandthecoloroftheprocessingpixelandcalculatesthecontrastbetweenthe