Image-Segmentation图像分割

EE4H,[e-mail]ThepurposeofimagesegmentationistopartitionanimageintomeaningfulregionswithrespecttoaparticularapplicationThesegmentationisbasedonmeasurementstakenfromtheimageandmightbegreylevel,colour,texture,depthormotion2IntroductiontoimagesegmentationUsuallyimagesegmentationisaninitialandvitalstepinaseriesofprocessesaimedatoverallimageunderstandingApplicationsofimagesegmentationincludeIdentifyingobjectsinasceneforobject-basedmeasurementssuchassizeandshapeIdentifyingobjectsinamovingsceneforobject-basedvideocompression(MPEG4)Identifyingobjectswhichareatdifferentdistancesfromasensorusingdepthmeasurementsfromalaserrangefinderenablingpathplanningforamobilerobots3IntroductiontoimagesegmentationExample1SegmentationbasedongreyscaleVerysimple‘model’ofgreyscaleleadstoinaccuraciesinobjectlabelling4IntroductiontoimagesegmentationExample2SegmentationbasedontextureEnablesobjectsurfaceswithvaryingpatternsofgreytobesegmented5Introductiontoimagesegmentation6IntroductiontoimagesegmentationExample3SegmentationbasedonmotionThemaindifficultyofmotionsegmentationisthatanintermediatestepisrequiredto(eitherimplicitlyorexplicitly)estimateanopticalflowfieldThesegmentationmustbebasedonthisestimateandnot,ingeneral,thetrueflow7Introductiontoimagesegmentation8IntroductiontoimagesegmentationExample3SegmentationbasedondepthThisexampleshowsarangeimage,obtainedwithalaserrangefinderAsegmentationbasedontherange(theobjectdistancefromthesensor)isusefulinguidingmobilerobots9Introductiontoimagesegmentation10OriginalimageRangeimageSegmentedimageGreylevelhistogram-basedsegmentationWewilllookattwoverysimpleimagesegmentationtechniquesthatarebasedonthegreylevelhistogramofanimageThresholdingClusteringWewilluseaverysimpleobject-backgroundtestimageWewillconsiderazero,lowandhighnoiseimage11Greylevelhistogram-basedsegmentation12NoisefreeLownoiseHighnoiseGreylevelhistogram-basedsegmentationHowdowecharacteriselownoiseandhighnoise?WecanconsiderthehistogramsofourimagesForthenoisefreeimage,itssimplytwospikesati=100,i=150Forthelownoiseimage,therearetwoclearpeakscentredoni=100,i=150Forthehighnoiseimage,thereisasinglepeak–twogreylevelpopulationscorrespondingtoobjectandbackgroundhavemerged13Greylevelhistogram-basedsegmentation140.00500.001000.001500.002000.002500.000.0050.00100.00150.00200.00250.00ih(i)NoisefreeLownoiseHighnoiseGreylevelhistogram-basedsegmentationWecandefinetheinputimagesignal-to-noiseratiointermsofthemeangreylevelvalueoftheobjectpixelsandbackgroundpixelsandtheadditivenoisestandarddeviation15SNbo/Greylevelhistogram-basedsegmentationForourtestimages:S/N(noisefree)=S/N(lownoise)=5S/N(lownoise)=216GreylevelthresholdingWecaneasilyunderstandsegmentationbasedonthresholdingbylookingatthehistogramofthelownoiseobject/backgroundimageThereisaclear‘valley’betweentotwopeaks17Greylevelthresholding180.00500.001000.001500.002000.002500.000.0050.00100.00150.00200.00250.00ih(i)BackgroundObjectTGreylevelthresholdingWecandefinethegreylevelthresholdingalgorithmasfollows:Ifthegreylevelofpixelp=TthenpixelpisanobjectpixelelsePixelpisabackgroundpixel19GreylevelthresholdingThissimplethresholdtestbegstheobviousquestionhowdowedeterminethethreshold?ManyapproachespossibleInteractivethresholdAdaptivethresholdMinimisationmethod20GreylevelthresholdingWewillconsiderindetailaminimisationmethodfordeterminingthethresholdMinimisationofthewithingroupvarianceRobotVision,Haralick&Shapiro,volume1,page2021GreylevelthresholdingIdealizedobject/backgroundimagehistogram220.00500.001000.001500.002000.002500.000.0050.00100.00150.00200.00250.00ih(i)TGreylevelthresholdingAnythresholdseparatesthehistograminto2groupswitheachgrouphavingitsownstatistics(mean,variance)ThehomogeneityofeachgroupismeasuredbythewithingroupvarianceTheoptimumthresholdisthatthresholdwhichminimizesthewithingroupvariancethusmaximizingthehomogeneityofeachgroup23GreylevelthresholdingLetgroupo(object)bethosepixelswithgreylevel=TLetgroupb(background)bethosepixelswithgreylevelTThepriorprobabilityofgroupoispo(T)Thepriorprobabilityofgroupbispb(T)24GreylevelthresholdingThefollowingexpressionscaneasilybederivedforpriorprobabilitiesofobjectandbackgroundwhereh(i)isthehistogramofanNpixelimage25pTPioiT()()0pTPibiT()()1255P(ihiN)()/GreylevelthresholdingThemeanandvarianceofeachgroupareasfollows:26oiToTiP(ipT())/()0biTbTiPipT()()/()1255ooiToTiTP(ipT220()())/()bbiTbTiTPipT221255()()()/()GreylevelthresholdingThewithingroupvarianceisdefinedas:WedeterminetheoptimumTbyminimizingthisexpressionwithrespecttoTOnlyrequires256comparisonsforand8-bitgreylevelimage27WoobbTTpTTpT222()()()()()Greylevelthresholding280.00500.001000.001500.002000.002500.000.0050.00100.00150.00200.00250.00ih(i)HistogramWithingroupvarianceToptGreylevelthresholdingWecanexaminetheperformanceofthisalgorithmonourlowandhighnoiseimageForthelownoisecase,itgivesanoptimum