机器视觉课件4

Course4.RegionsDef.ofRegion:Asetofconnectedimagepixels,withsimilarproperties.•similargraylevel•similarcolor•similartexturesoon……Regionmaybe•separated•adjacent•overlapped•includedImportantprincipleofregionprocess:——homogeneitypredicate(valuesimilarity)——specialproximity1.RegionRepresentation1)Arrayrepresentation:GivenaregionRinimageS,ItsarrayrepresentationAis:—sameasbinaryimagerepresentation—oftenbeusedasamaskoverintensityimagetoselecttheregionforproperprocessing.otherwiseRjiifjiaSjijiaA0],[1],[,],[],[2)PyramidsrepresentationPyramidrepresentationofaNxNimage(N=2k)containsklayersofreducedimage.Eachpixelinanupperlayerimagecorrespondsto4pixelsinthelowerlayerimage,i.e.,Pixelsvalueofupperlayerimagecanbeobtainedby——averagingthecorrespondingpixelsoflowerlayerimage,or——subsampling(e.g.chooseupper-leftpixelvalue)114422NNNNNN3)QuadTree(Quad=4)Children:ifaregionSispartitionedinto4subregion,inquadtree,eachsubregionnodeiscallachildofthenodeoftheoriginalregion.nodestwobetweenrelationslinksregionofpropertynodesTree——1234S1234Colorofnode:Algorithm:a)Setanimageasaregion,andrepresentsitasanodeinquadtree(0-level)b)Splittheregioninto4subregions,andrepresentthemin4childnodes.c)Checkthecolorofeachchildnode:pixelsandbothhavegrayregiontheinarepixelsallblackregiontheinarepixelsallwhitecolor10.1.0.——ifcolorisblackorwhite,setnodeas“leafnode”,nofurthersplitting.——ifcolorisgray,gotostepb).d)Stopifallnodesareeitherblackorwhite.4)RegionAdjacencyGraph:——Emphasizetheadjacencyofregions——node:region——link:commonboundarybetweenregions.5)PictureTree:——Emphasizeregioninclusionwithinanotherregion.——Recursivelysplitanimageintoregions,andtheprocessstopswhennoregioncanbefurthersplitted(homogeneitypredicate)2.RegionSegmentationSupposeanimageAiswellsegmentedintonregionsRi,i=1,2,…,n,theremustbe:WhereP(.)ishomogeneitypredicatemeasurement.FalseRRPTrueRPARjiinii)()(11)Thresholding:(1)P-TileThresholingIfapre-knowledgeofobjectsizeSisgiven,theareapercentageoftheobjectinimageAcanbegotbyP=S/A.Thus,wesetthevalueofthresholdatTsuchthatwhere0TLandcanbeobtainedfromimage’shistograph.TdpP0)(Lp01)()(p(2)Optimalthresholding:SupposeanimagecontainonlytwoprincipalbrightnessregionsandbothobeyGaussiandistributioninintensity.Darkregion:Brightregion:wesetpartitionthresholdatT,errorprobabilityofobjectbeingsegmentedasbackground:21212)(111)(σμZeσ2πZp22222σ)μ(Z22eσ2π1(Z)p21TdpTEττ)()(21Errorprobabilityofbackgroundbeingsegmentedasobject:IftheprioriprobabilitiesofbackgroundandobjectareP1andP2respectively,theoverallerrorprobabilityis:TshouldbechosensuchthatT1dτ(τpTE))(2)()()(2112TEPTEPTEmin)(TE,0)(TTEFromGetApplyingGaussiandistributions,get:Where:)()(2211TpPTpP02CBTAT)ln(2)(221122221212222212222112221PPCBAOnecansolveforTInthecaseWecangetInmostapplication,wejustsimplythresholdanimageatintensityTwherehistographhasavalley,whichcanapproximateoptimalthresholding.22221)ln(21221221PPT(3)AdaptiveThresholding:ThebasicconceptistochoosethresholdTlocallytoagainstunevendistributionofimageintensity(causedbyunevenilluminationinscene).•Partitionanimageintoseveralregions.•Ineachregion,choosethresholdTatmajorvalleyofithistograph,andperformthreholdingineachregion.(4)VariableThresholding:•Thresholdingagainstunevenintensitydistribution(unevenilluminationinscene).•Approximateintensitydistributionofimagebackgroundbyaplaneoraquadraticsurface.•Performthresholdingrelatingtothebaselevelofintensityatthesurface.2)RegionMerging•Thresholdinggeneratestomanyextraregionsthatwedonotwant.•Mergeoperationcombinesregionsthatareconsideredsimilar.Algorithm:a)Forminitialregionsintheimageusingthresholding,followedbyclusteringlabeling(findconnectedcomponents).b)Makeanadjacencygraphfortheimage.c)Foreachregionintheimage,dothefollowingsteps:——Consideritadjacentregions,andcheckwhethertheyaresimilar.——Fortheregionthataresimilar,mergethemandmodifytheadjacencygraph.d)Repeatstepc)untilnoregionaremerged.Criteriaformerge:1)Meanintensityofregion2)Probabilitydistributionofintensity3)Weaknessofboundary(edge):ABABIfmergeregionsAandBWhere—weakedgelength—commonedgelengthT—threshold,e.g.T=0.75Howtomeasuretheweaknessofanedge?T—presetvaluei.e.,thegradientofimageintensity.τSwwsTI3)RegionSpliting:——ifsomepropertyofaregionisnotconstant,theregionshouldbesplit——difficulties:howtomeasure“properties”howtofindsplittingboundary?Algorithm(splitandmerge):a)SetthewholeimageasaregionRb)PickaregionR,calculateintensityvarianceintheregion,ifislarge,splittheregionintofoursub-regions.c)Considertwoormoreneighboringregions,Let.CalculateintensityvarianceinR.Ifissmall,mergethenregionsintoasingleregion.d)Repeatstepb)andstepc)untilnoregioncanbesplitandmerged.RegionGrowing:——Aggressivelyincreasethesizeofaseedregionbycheckingthehomogeneitypredicateofitsneighboringpixels.nRRRR21nRRR,,,21——Homogeneitypredicatecanbebasedonanycharacteristicoftheregion,e.g.,intensity,variance,color,texture,…Example:Regiongrowingbasedonsurfacefitting:•Fittingintensityofare