基于张量分析的不完备图像修复研究

分类号学号2005611810203学校代码10487密级硕士学位论文基于张量分析的不完备图像修复研究学位申请人：黄应学科专业：通信与信息系统指导教师：田岩副教授答辩日期：2007年2月24日ADissertationSubmittedinPartialFulfillmentoftheRequirementsfortheDegreeofMasterofEngineeringStudyofIncompleteImageRepairingTechnologyUsingTensorAnalysisCandidate:HuangYingMajor:Communication&InformationSystemSupervisor:AssociateProf.TianYanHuazhongUniversityofScienceandTechnologyWuhan,Hubei430074,P.R.ChinaFebruary,2007独创性声明本人声明所呈交的学位论文是我个人在导师指导下进行的研究工作及取得的研究成果。尽我所知，除文中已经标明引用的内容外，本论文不包含任何其他个人或集体已经发表或撰写过的研究成果。对本文的研究做出贡献的个人和集体，均已在文中以明确方式标明。本人完全意识到，本声明的法律结果由本人承担。学位论文作者签名：日期：年月日学位论文版权使用授权书本学位论文作者完全了解学校有关保留、使用学位论文的规定，即：学校有权保留并向国家有关部门或机构送交论文的复印件和电子版，允许论文被查阅和借阅。本人授权华中科技大学可以将本学位论文的全部或部分内容编入有关数据库进行检索，可以采用影印、缩印或扫描等复制手段保存和汇编本学位论文。保密□，在______年解密后适用本授权数。本论文属于不保密□。（请在以上方框内打“√”）学位论文作者签名：指导教师签名：日期：年月日日期：年月日华中科技大学硕士学位论文I摘要在许多实际问题中，图像可能在获取、传输或者保存的过程中产生不可避免的缺陷，如医学扫描图像中的三维器官和病灶、遥感影像中的缺损现象、珍贵影像资料保存不善带来的划痕等，这类有缺陷的数据就称作不完备数据。不完备图像的处理和识别目前是一个崭新的研究方向。其方法之一，是采用有效的修补或补全技术，将不完备图像进行修复，得到缺失的部分的数据，进而进行后续的相关处理和识别。现有的图像修复技术主要可以分为两类：一类是基于几何图像模型的图像修补(inpainting)技术，该技术特别适用于修复图像中的小尺度缺损；另一类是基于纹理合成的图像补全(completion)技术，该技术对于填充图像中大的丢失块有较好的效果。本文工作的重点是研究对不完备图像数据完备化的方法。考虑到张量分析技术能够提供图像的几何结构信息，本文研究了基于张量分析的不完备数据的分析和修复等方法。论文首先详细介绍了图像的张量投票及其特征提取方法，并对不同的应用提出了两种改进的张量投票算法。在此基础上，本文分别提出了针对小区域缺失的自适应图像修复算法和针对大区域缺失的修复算法。在小区域缺失的修复算法中，建立了一种基于张量特征的优先权，用迭代的办法首先修复具有最大优先权的窗口，由于是自动选取窗口并根据窗口大小的不同而选择不同的阈值，这种方法具有更好的适用性。大区域修复算法则采用了一种结构修复和纹理修复的方法，首先补全图像缺失的结构信息，然后从结构和纹理上对图像进行修复，由于引入了对图像结构的推测，本文的算法能够取得更好的视觉效果和图像客观评价指标。实验表明本文的算法与传统方法相比，具有更好的图像修复效果。本文对不完备数据修复的工作也将拓展传统数字图像处理、计算机视觉、模式识别等方法的处理和应用范围。关键词：不完备数据;张量投票;图像修复;纹理合成华中科技大学硕士学位论文IIAbstractInmanypracticalproblems,theimagesmayinevitablyhaveflawduringtheimageacquisition,transmissionorpreservationprocesses,suchasthethree-dimensionalmedicalscanningimagesofpathologicalorgans,theremotesensingimageswithdefects,andthescratchedpreciousimagespreservedunderpoorcondition.Thisthesistakesthosedefectiveimagesastheincompleteimagedata.Incompleteimageprocessingandrecognitionnowisanewdirectionintheresearches.Oneofthesolutionsistomakeaneffectiverepairorcompletiontotheincompletedata,andthenimplementrelevantprocessingorrecognition.Thecurrentimagerepairingtechnologiescanbedividedintotwocategories.Oneisthegeometricmodelbasedimageinpaintingtechnology,whichisparticularlyapplicabletorepairthesmall-scaleimagesdefects.Theotheristhetexturesynthesisbasedimagefillingorcompletiontechnology,whichachievesabetterresultinfillingtheimageswithlarge-areamissingpieces.Thefocusofthispaperistotakeresearchoneffectiveimagerepairingmethodsfromincompletedataset.Consideringthattensoranalysisisaveryusefulmethodtoconferthegeometricstructureinformationfromsparseornoisedimages.Thisthesisexpectstoconductanovelincompletedataanalysisandrepairingmethodinthevirtueoftensoranalysis.Atthefirst,tensorvotinganditsfeatureextractionmethodswouldbepresented.Thepaperalsomakesafewimprovementstotheexistingtensorvotingalgorithm.Onthisbasis,weproposeimagerepairingalgorithmsfordeletionofsmallregionalandlargeregionaldamagerespectively.Inthesmallarearepairingalgorithm,webuildaprioritybasedonthetensorfeaturesandthenuseaniterativemethodtorepairtheimagewindowwithgreatestpriority.Asthesizeofthewindowisautomaticallyselectedanddifferentthresholdsareappliedaccordingly,thismethodhasbetterapplicabilityinthisfield.Inthelargearearepairingalgorithm,ourmethodincludingbothstructuralandtexturerepairingprocesses,thatisfirsttorepairthemissedstructurepropertyoftheentireimage,andthento华中科技大学硕士学位论文IIIrepairthestructureandtexturewithin.Becauseoftheinferenceofthemissedstructuralinformation,ouralgorithmachievesabettervisualeffectandobjectiveevaluationcriterion.Atlast,experimentsshowthatcomparingwithtraditionalalgorithms,andourmethodhasabetterresultinrepairingincompleteimages.Therefore,theworkinthispaperwouldextendtheapplicabilityofconventionaltechnologiesindigitalimageprocessing,computervisionandpatternrecognitionareas.Keywords:IncompleteData;TensorVote;ImageRepairing;TextureSynthesis华中科技大学硕士学位论文IV目录摘要···················································································1Abstract·················································································II1绪论1.1引言···············································································(1)1.2不完备数据修复································································(3)1.3张量投票技术···································································(5)1.4论文的主要工作及内容安排·················································(7)2图像的张量投票2.1引言···············································································(8)2.2张量和张量分解算法························································(10)2.3图像的张量表达和张量投票算法·········································(11)2.4图像结构信息的提取························································(14)2.5改进的张量投票方法························································(15)2.6实验结果和分析······························································(16)2.7小结·············································································(19)3基于张量分析的自适应小区域缺失图像修复算法3.1图