搜索
2.3.4.5.6.7.8-11.12.13.“superpixels”RenandMalik,2003BerkeleysegmentationdatabaseimagehumansegmentationGestalt••,1,,0,,••—p1962Doylep-p%—1966Prewitt—1.K=0TK2.TKTKTATKTB3.TK1=(TA+TB)/2TK1TK4.K=K+1;5.2-5TKTK1TK—Otsu1979OtsuOtsu—OtsuOtsu•tw0u0w1u1u=w0u0+w1u1•ttg=w0(u0-u)2+w1(u1-u)2t•Otsug=w0w1(u0-u1)2—Otsu—WatershedWatershedVincentandSoille,19911.2.3.4.—Watershed1.2.h(FIFO)3.FIFO4.35.(3D)(3D)—Watershed—Watershed12()3()4ActiveContourModelActiveContourModel—SnakesSnakesActiveContourModelKass,Witkin,andTerzopoulos1988,—Snakes,Etotalf(i)—SnakesSnakes832—SnakesSnakesballooningforce—Snakes+f(s)—CondensationSnakeKalmanSankes—Condensation—CondensationParticleFiltering—CondensationIsardBlake1998CONDENSATION(CONditionalDENSitypropagATION)IsardandBlake1998—Condensation—Condensation—LevelSetsSnakesCondensationMalladi,Sethian,andVemuri,1995LevelSets,ZeroLevelSet,x-yNN+1NN+1NSnake:ParametricActiveContourLevelSets:GeometricActiveContours—LevelSetsCremers,Rousson,andDeriche,2007Clustering•KK-means•MeanShift•KK-meansparametric•MeanShiftnon-parametric—K-means(R,G,B)(R,G,B,X,Y)(L*u*v)L*u*v—K-meansK-means1.k2.kμ1,μ2,…,μk3.xiciargmin4.jμj5.3-5μ1,μ2,…,μk—K-meansK-meansk=2—K-meansK-meansExpectationMaximization,EMK-means,Σxi,,Σ|,Σp(x),Σ,…,,Σ—K-meansEM•E-stepZ|,Σ∑1zikxik•M-step∑1Σ1—K-means—K-meansK-means12K-means1Memory-intensive2k345—MeanShiftMeanShiftK-meansComaniciuandMeer,2002—MeanShiftMeanShiftL*uL*u—MeanShiftL*uL*u—MeanShiftxik()Parzenwindowmax 0,1exp12EpanechnikovkernelGaussiankernel—MeanShiftf(x)Meanshift,∑∑meanshift—MeanShiftSearchwindowCenterofmassMeanShiftvector—MeanShiftSearchwindowCenterofmassMeanShiftvector—MeanShiftSearchwindowCenterofmassMeanShiftvector—MeanShiftSearchwindowCenterofmassMeanShiftvectorSearchwindowCenterofmass—MeanShiftMeanShiftvectorSearchwindowCenterofmass—MeanShiftMeanShiftvectorSearchwindowCenterofmass—MeanShift—MeanShiftMeanshift1.2.3.Meanshift4.L*u159meanshift—MeanShiftComaniciuandMeer,2002—MeanShiftMeanshift—MeanShiftMeanshift—MeanShiftMeanshift1234Meanshift12—NormalizedcutsABcutcut,∈,∈minimumcut—Normalizedcuts,∈,∈exp —NormalizedcutsminimumcutcutNormalizedcutCut1Cut2—NormalizedcutsNormalizedcut,,,,,Ncut(A,B)—NormalizedcutsNormalizedcutNcut(A,B)cutNormalizedcutNP-hardNormalizedcutW=[ωij]D=diag(d(1),…,d(N))d(i)Wi(D-W)y=λDyNcut—Normalizedcuts3232—NormalizedcutsWDShiandMalik,2000—GraphcutsMaximumflowMinimumcut—GraphcutsST—GraphcutsLL0-101LE(L)=Ed(L)+λEs(L)/—GraphcutsE(L)=Ed(L)+λEs(L),,,,,,,,∞ ,,, if ,,otherwise ,,0,,1(x,y)(x,y)—Graphcuts,,0,,1(x,y)(x,y),,0,,1—GraphcutsE(L)=Ed(L)+λEs(L),ωpq=10.0ωpq=0.1ωpqpq—GraphcutsE(L)=Ed(L)+λEs(L)maxflow/mincutts (“foreground”)min cut(“background”),,1,,0mincuts-t—GraphcutsGraphcutsRother,Kolmogorov,andBlake,2004——p———Otsu—Watershed—Snake—Condensation—LevelSets—K-means—Meanshift—Normalizedcuts—Graphcuts—•Kass,M.,Witkin,A.,andTerzopoulos,D.Snakes:Activecontourmodels.IJCV,1(4):321–331,1988.•Isard,M.andBlake,A.CONDENSATION—conditionaldensitypropagationforvisualtracking.IJCV,29(1):5–28,1998.•Comaniciu,D.andMeer,P.Meanshift:Arobustapproachtowardfeaturespaceanalysis.IEEET-PAMI,24(5):603–619,2002.•Shi,J.andMalik,J.Normalizedcutsandimagesegmentation.IEEET-PAMI,8(22):888–905,2000.•Boykov,Y.andKolmogorov,V.Anexperimentalcomparisonofmin-cut/max-flowalgorithmsforenergyminimizationinvision.IEEET-PAMI,26(9):1124–1137,2004.