资源描述：
《Efficient Laplacian-based matting Using Gaussian KD-tree.doc》由会员上传分享，免费在线阅读，更多相关内容在教育资源-天天文库。
1、EfficientLaplacian-basedmattingUsingGaussianKD-treeAbstract:Weproposedaunifiedacceleratingmethodforlaplacian-basedmattingusingGaussianKD-tree.IntroductionRelatedwork本文的主要思想，流程（1）构建KD树（对图像进行采样）（GPU加速）（2）交互（前景点，背景点）（3）将交互的点集扩散到KD树上（GPU加速）（4）构建拉普拉斯矩阵时搜索最近领域（GPU加速）（5）求解拉普拉斯矩阵（GPU加速）（6）在低
2、分辨率matting的结果插值回原图像（GPU加速）在如下三个方面需要利用到最近领域搜索，都可利用GPU加速（1）将交互的点集扩散到KD树上（GPU加速）（2）构建拉普拉斯矩阵时搜索最近领域（GPU加速）（3）在低分辨率matting的结果插值回原图像（GPU加速）主要应用（1）图像的快速去雾（2）视频的快速去雾需要解决的问题（1）GMM模型用于采样及交互（2）更好的采样，参考，OptimizedColorSamplingforRobustMatting（CVPR07）（3）再修改laplacian-basedmatting，提出一个Bilateralweigh
3、tedclosed-formmatting现有方法的缺点CVPR2010.计算量大，耗内存，需要三分图，不适合videomattingClosed-formMattingBothPoissonmattingandGeodesicmattingapproachesinvolveestimatingforegroundandbackgroundcolors/distributionstosomeextent,whichmaysignificantlylowertheirperformancewhentheestimationsarenotaccurate.There
4、centlyproposedclosed-formmattingapproach[Levin06]avoidsthislimitationbyexplicitlyderivingacostfunctionfromlocalsmoothnessassumptionsonforegroundandbackgroundcolorsFandB,andshowthatintheresultingexpressionitispossibletoanalyticallyeliminateFandB,yieldingaquadraticcostfunctionin,whichc
5、anbeeasilysolvedasasparselinearsystemofequations.TheunderlyingassumptionmadeinthisapproachisthateachFandBisalinearmixtureoftwocolorsoverasmallwindow(typically3×3or5×5)aroundeachpixel,whichisreferredtoasthecolorlinemodel.Itisshownthatunderthisassumption,alphavaluesinasmallwindowwcanbe
6、expressedas,,wherereferstocolorchannels,andandareconstantsinthewindow.ThemattingcostfunctionisthendefinedasFurthermore,andcanbeeliminatedfromthecostfunction,yieldingaquadraticcostinthealone:,whereisanN×Nmatrix,whose(i,j)-thelementis:whereisa3×3covariancematrix,isa3×1meanvectoroftheco
7、lorsinawindow,andisthe3×3identitymatrix.ThematrixL,whichiscalledmattingLaplacian,isthemostimportantanalyticresultfromthisapproach.Theoptimalalphavaluesarethencomputedas,s.t.whichisessentiallyaproblemofminimizingaquadraticerrorscore,thuscanbesolvedbyoneofthelinearsystemsolverswhichwil
8、lbediscussed
