计算机研究与发展JournalofComputerResearchandDevelopmentlSSN1000—1239/CN11—1777/TP42(7):l179~ll84,2005Snake模型与深度凹陷区域的分割王元全汤敏王平安夏德深徐晔。(南京理工大学计算机科学与工程系南京2l0094)(香港中文大学计算机科学与工程系香港)(江苏油田南京计算中心南京2l0046)(fitl00@sohu.corn)ResearchonBoundaryConcavitiesSegmentationviaSnakeModelsWangYuanquan,TangMin,PhengAnnHeng,XiaDeshen,andXuYe。(Departmentof’ComputerScience&Engineering,NanjingUniversityof&ienceandTechnology.Nanjing210094)(DepartmentofComputers(ience&Engineering,TheChineseUniversity,HongKong,HongKg)(NanjingGeophysicalProspectingResearchInstituteofJiangsuOilfield,Nanjing210046)AbstractSnakemodelsareextensivelyusedfromitsdebutinimageprocessingandmotiontracking,butitspoorconvergenceonconcaveboundaryisahandicapforobjectlocation.Although,theGVFSnakemodelshowshighperformanceforthisproblem,butitsuffersfromcostlycomputationbyvirtualofPDE’sandanotherso—calledcriticalpointproblemfortheinitialcontourselection.InordertoimprovetheperformanceofthetraditionalSnakemodelforconcavitysegmentation,anewexternalforcebasedonthelocalcurvatureofthediscretecontourandatwo—stageSnake—basedalgorithmareproposed.Thelocalcurvatureofthediscretecontour,whichcharacterizesthebendingofacontourassociatedwithadirection,isdefinedusingthecenteroftheinscribedcircleofthetrianglederivedfromthreeconseeutivecontournodes.ThefirststageofthenewmethodisatraditionalSnake,andinthesecondstagethenewforcewoulddrivethecontourintotheconcaveregion.ThisnewforcecanalsobegeneralizedtoenlargethecapturerangeoftheSnakemode1.Inthiscase,itcanbeconsideredasageneralizationoftheballoonforce.Inordertoovercomethedifficultyofdeterminingthemagnitude,themagnitudeissettobesmallandthegradient—basedforceisfirstusedasresistance;whenthecontourisconverged,thegradient—basedforceswervestoattractthecontour.Generalizedinthisway,thecapturerangeisenlargedandthereisnocriticalpointproblem.Theexperimentalresultsvalidatetheperformanceofthismethod.KeywordsSnakemodels;GVFSnakemodel;concavitysegmentation摘要在基于Snake模型的图像分割中,深度凹陷区域的分割是一个难点.尽管GVFSnake模型极大地改善了这个问题,但它需要事先求解一个偏微分方程组,增大了计算量;同时,GVFSnake模型在初始化时还存在一个“临界点”问题.探讨了深度凹陷区域的分割,用离散轮廓上顺序3点所成三角形内切圆圆心来定义离散轮廓的曲率,该曲率既能反映轮廓的弯曲程度,又具有合理的方向.基于该曲率定义了曲率外力项,构造了基于Snake模型的两阶段算法,先采用传统Snake模型使离散轮廓逼近目标边缘,然后在曲率外力的作用下使离散轮廓进入目标的凹陷区.曲率外力项的引入能较好地解决深度凹陷区域的分割问题,也可以扩展该外力项来扩大Snake模型的捕捉范围.实验结果表明该方法是有效的.关键词Snake模型;GVFSnake模型:深度凹陷分割中图法分类号TP39141收稿日期:20030924;修回日期:20040510基金项目:香港特别行政区政府研究资助硒项目(cuHK4l80/0lE。CUHKI/00C)维普资讯http://www.cqvip.comll80计算机研究与发展2005,42(7)1引言传统的图像理解方法认为轮廓提取、运动跟踪等只能依赖于来自图像的知识,然后对这些知识选取一个合适的描述方式,这是一个自底向上的过程.但Snake模型的出现改变了这种状况llj,它有两个主要特点:一方面它将目标轮廓的初始估计、曲线性质、图像信息和先验知识融于一个统一的过程中;另一方面,它在分割目标边缘的过程中寻找能量泛函的局部极小点,而非全局极小点.基于Snake的方法是一个自顶向下的过程.Snake轮廓相当于一条弹性曲线,该曲线在内部能量...