基于非参数模型的点集匹配算法研究-控制科学与工程专业论文.docx

遮挡、离群点、旋转、和尺度变换的鲁棒性，在大量公共数据集上与处于领先水平的算法作定性和定量的对比，实验结果表明RPM-?2? 能得到更精确的点集或图像配准结果。然后，针对计算机图形学与图像处理领域的传统课题――非刚性图像变形问题，提出了一种基于移动正则化最小二乘的非刚性图像变形算法 MRLS。该算法可生成生动直观的变形效果，且计算效率极高可满足实时性要求，从而不仅为图像配准算法定量评估提供必要的实验数据，而且在动画制作、形体塑造、医学图像分析等其它应用中也具有潜在价值。最后，针对非刚性点集配准问题提出了一种基于保持点集全局与局部结构特征的非刚性点集配准算法PR-GLS，其采用高斯混合模型进行建模，结合软指派策略将点集对应关系和变换函数放在统一框架下进行估计，从而改进配准效果。相比于RPM-?2?和其它处于领先水平的点集配准算法，PR-GLS在数据退化程度较为严重 时能大大提升配准精度。关键词：点集匹配，非参数模型，图像配准，向量场，离群点，非刚性形变，稀疏近似AbstractPointmatching,whichaimstoestablishcorrespondencebetweentwogivenpointsets,isafundamentalandcriticalproblemintheareasofcomputervisionandpatternrecogni-tion.Ithasrecentlyreceivedalotofattentionsinawiderangeofapplicationsincludingstereomatching,objectrecognitionandtracking,medicalimageanalysis,remotesensingimage processing,etc.Pointmatching isessentially anNPCcomplicatecombinatorial opti-mizationproblemwithlargecomputation;andduetotheexistenceofnoise,outliers,aswellasunknowndeformations,theperformancesofpointmatchingalgorithmsareoftenbadlydegraded,whichlimitstheirsapplicabilitiesinpracticalengineeringproblems.Therefore, researchinguniversal,efficientandrobustpointmatchingmethodshashightheoreticsignif-icanceandpracticalvalues.Inordertoefficientlyhandletheeffectsofnoise,outliersandnon-rigiddeformations inpointmatching,thisdissertationaimstoresearchrobustpointmatchingmethodsforgeneraltasks.Tothisend,inthedissertation,severalnovelalgorithmsareproposedbasedonnon-parametricmodel,andareappliedtoimagefeaturepointmatching,non-rigidpointsetregistration, as well asnon-rigid image registration.The experimentalresults on various2Dand3Dcasesshowthattheproposedmethodsoutperformsmanystate-of-the-artmethods.Themainworksandcontributionsofthisdissertationareasfollows.First,theproblemofvectorfieldinterpolationiscarefullystudiedintheframeworkofregularizationandtherepresentertheorem.ArobustvectorfieldinterpolationalgorithmVFCisproposed,whichisabletohandleaverylargenumberofoutliers.Toimprovetheefficiencyofkernel-basedmethods, thedissertationintroducesasp