基于目标基于目标区域

《基于目标基于目标区域》由会员上传分享，免费在线阅读，更多相关内容在工程资料-天天文库。

学校编码学校编码：学校编码：：：10384103841038410384分类号分类号分类号密级密级密级学号学号：学号：：：2302010115302302010115302302010115306023020101153060UDCUDCUDCUDC硕士士士士学学学学位位位位论论论论文文基于目标区域的图像检索技术研究ImageRetrievalBasedontheObjectRegions郝昕疌指导教师姓名：雷蕴奇教授专业名称：计算机应用技术论文提交日期：2013年5月论文答辩时间：2013年6月学位授予日期：2013年月厦门大学博硕士论文摘要库答辩委员会主席：评阅人：2013年5月 厦门大学博硕士论文摘要库 厦门大学学位论文原创性声明本人呈交的学位论文是本人在导师指导下,独立完成的研究成果。本人在论文写作中参考其他个人或集体已经发表的研究成果，均在文中以适当方式明确标明，并符合法律规范和《厦门大学研究生学术活动规范（试行）》。另外，该学位论文为（计算机视觉）课题（组）的研究成果，获得（计算机视觉）课题（组）经费或实验室的资助，在（雷蕴奇）实验室完成。（请在以上括号内填写课题或课题组负责人或实验室名称，未有此项声明内容的，可以不作特别声明。）声明人（签名）：年月日厦门大学博硕士论文摘要库 厦门大学博硕士论文摘要库 厦门大学学位论文著作权使用声明本人同意厦门大学根据《中华人民共和国学位条例暂行实施办法》等规定保留和使用此学位论文，并向主管部门或其指定机构送交学位论文（包括纸质版和电子版），允许学位论文进入厦门大学图书馆及其数据库被查阅、借阅。本人同意厦门大学将学位论文加入全国博士、硕士学位论文共建单位数据库进行检索，将学位论文的标题和摘要汇编出版，采用影印、缩印或者其它方式合理复制学位论文。本学位论文属于：（）1.经厦门大学保密委员会审查核定的保密学位论文，于年月日解密，解密后适用上述授权。（）2.不保密，适用上述授权。（请在以上相应括号内打“√”或填上相应内容。保密学位论文应是已经厦门大学保密委员会审定过的学位论文，未经厦门大学保密委员会审定的学位论文均为公开学位论文。此声明栏不填写的，默认为公开学位论文，均适用上述授权。）厦门大学博硕士论文摘要库声明人（签名）：年月日3 厦门大学博硕士论文摘要库 摘要经过二十多年的发展，解决图像视觉特征与语义特征之间的差异问题已成为基于内容图像检索的研究热点。本文认为获取图像中不同目标的类别和相对位置信息是构建图像“语义特征”的重要基础。为此，深入研究了图像分类和图像多目标区域提取这两个“打好基础”的关键技术，主要研究工作包括：1、从颜色、纹理、形状和伪语义信息四个方面研究了图像的低层特征。通过在Corel-1k图像库上的分类实验，分析、讨论了各种低层特征对不同类型图像的描述能力，为特征融合和图像分类积累丰富的先验知识。2、设计实现了三种特征融合方法。其中前两种是“对症下药”的特征级融合方法，该类方法针对图像类型或者图像不同部分的特点选取适合的图像低层特征，再拼接这些特征作为融合特征。第三种是“机器学习”的决策级融合方法。该方法是在图像分类前，选择多种图像低层特征和对应的SVM核函数，然后利用SVM多重核函数学习方法不断调整各种特征的权重以训练出分类效果最优的分类器。本文分别使用三种方法对Corel-1k图像库进行了分类实验，结果表明根据图像特点选择互补性好的特征进行融合可以获得比单一特征更优的图像分类效果。3、提取图像中多个目标区域。本文首先采用改进的光谱残差法、边缘检测和滑动窗口这三种方法分别提取出图像中候选目标区域，然后利用颜色对比度模型剔除非目标的候选区域，再利用一种新的超像素块模型构造目标区域的评价函数，最后选取评价值较高的候选区域作为最终的目标区域。在Corel-1k图像库上将本文的目标区域提取效果与目前几种先进的算法作了效果对比，并在MSRC-v2图像库上根据手工标注的目标区域进一步测试了本文算法的性能，结果表明本文方法可以有效提取图像中不同尺度的多个目标区域，但是仍有一些不足需要改进。最后，提出了一个三层的图像多目标区域提取框架。前两层结构可以不断融入新厦门大学博硕士论文摘要库方法来提高目标区域提取能力，最后一层结构可自行定义以提取出满足特定需求的目标区域，为构建语义信息打下坚实基础。本文使用C++,OpenCV,Matlab和SQLServer数据库构建了以上述算法为基础的图像检索系统。关键词关键词：关键词：：：特征融合；SVM多核学习；多目标区域提取；颜色对比度；超像素I 厦门大学博硕士论文摘要库 AbstractBridgingthedifferencesbetweenvisualfeaturesandsemanticfeatureshasbecomingaresearchhotspotinthefieldofcontent-basedimageretrievalaftertwodecadesofdevelopment.Thisthesisarguesthataccesstovariousobjectsandtheirrelativepositioninformationisthesignificantfoundationofconstructingsemanticfeatures.Tothisend,thisthesisstudiesimageclassificationandmulti-objectregionsextraction,whicharetwocrucialtechniquesoflayingthefoundationofsemanticfeatures.Themainresearchworksareasfollows.1.Fourkindsoflow-levelfeatures-color,texture,shapeandpseudo-semanticinformation-havebeenstudied.BytakingimageclassificationexperimentsonCorel-1kimagedataset,theabilityofdifferentkindsoflow-levelfeaturestodescribevariouskindsofimagesisanalyzedanddiscussed,throughwhichawealthofpriorknowledgeisaccumulatedforthelaterfeaturefusionandimageclassification.2.Threefeaturefusionmethodshavebeendesignedandimplemented.Thefirsttwoarefeature-levelfusionmethods.Thesemethodsholdtheideaofsplicingseverallow-levelfeatures,whicharechosenaccordingtothecharacteristicofthewholeorpartsoftheimage,asonefusionfeature.Thethirdisadecision-levelfusionmethodusingmachinelearningtheory.Atfirst,avarietyoflow-levelimagefeaturesandcorrespondingSVMkernelfunctionsarechosen.ThentheclassifierwillnotstopadjustingtheweightsoffeaturesunderthecontrolofmultiplekernelSVMlearningmethoduntilthebestclassificationresultisachieved.AllthethreemethodsweretestedonCorel-1kdataset,theresultsshowthatfu厦门大学博硕士论文摘要库sionfeaturewithexcellentcomplementarityoutperformssinglefeatureinimageclassification.3.Extractingapluralityofobjectregions.Firstly,improvedspectralresidualmethod,edgedetectionmethodandslidingwindowmethodareappliedtoextractcandidateobjectregions,respectively.Secondly,amodelbasedonColorContrastisappliedtoexcludenon-objectcandidateregions.Thirdly,amodelutilizingsuperpixelsIII isappliedtocreateanevaluationfunction,whichisusedtojudgethecandidateobjectregions.Finally,candidateregionswithhigherevaluationvaluearemorelikelytobeselectedasfinalobjectregions.TheresultofextractingobjectregionsonCorel-1kdatasetbythisthesisiscomparedwiththeresultsofsomestate-of-the-artmethods.Moreover,furtherperformancetestistakenbycomparingtheresultofthisthesiswiththegroundtruthonMSRC-v2dataset.Theexperimentresultsshowthemethodofthisthesisisgoodatextractingobjectregionsundervariousscales,whereastheshortcomingsremainalot.Atlast,athree-tierframeworkformulti-objectregionsextractionispresented.Theformertwo-tierstructuresareabletointegrateintonewmethodstoimprovetheobjectregionsextractioncapability.Thethirdtiercanbedefinedfreelytomeetthespecificneedsofobjectregions,whichhelpstobuildasolidfoundationofsemanticinformationstructure.AnimageretrievalsystembasedonthealgorithmsmentionedaboveisimplementedusingC++,OpenCV,MatlabandSQLServer2000.Keywords:FeatureFusion;MultipleKernelLearningSVM;Multi-objectregionsextraction;ColorContrast;SuperPixels厦门大学博硕士论文摘要库IV 目录摘要..............................................................................................................IAbstract....................................................................................................III目录.............................................................................................................VContents....................................................................................................IX第一章绪论................................................................................................11.1选题研究背景与意义.......................................................................................11.1.1选题研究背景.........................................................................................11.1.2选题研究意义.........................................................................................21.2图像检索研究现状与热点...............................................................................21.2.1图像检索研究现状.................................................................................21.2.2图像检索研究热点.................................................................................31.3课题组以往工作总结.......................................................................................61.4本文主要内容及组织结构...............................................................................7第二章第二章第二章图像低层特征的提取与对比图像低层特征的提取与对比.......................................................92.1引言...................................................................................................................92.2颜色特征...........................................................................................................92.2.1颜色空间.................................................................................................92.2.2颜色矩...................................................................................................10厦门大学博硕士论文摘要库2.2.3颜色直方图...........................................................................................132.3纹理特征.........................................................................................................152.3.1纹理的定义...........................................................................................152.3.2纹理特征提取方法................................................................................152.3.3不同纹理特征分类实验.......................................................................212.4形状特征.........................................................................................................222.4.1形状特征提取方法...............................................................................22V 基于目标区域的图像检索技术研究2.4.2不同形状特征分类实验.......................................................................252.5伪语义信息特征.............................................................................................252.5.1伪语义信息特征的定义.......................................................................252.5.2伪语义信息特征提取方法...................................................................262.5.3伪语义信息特征特征分类实验...........................................................272.6本章小结.........................................................................................................27第三章第三章第三章图像特征融合图像特征融合..............................................................................293.1引言.................................................................................................................293.2图像特征的特征级融合.................................................................................293.2.1拼接特征融合.......................................................................................293.2.2基于图像显著区域的特征融合...........................................................323.3图像特征的决策级融合.................................................................................343.3.1SVM多重核函数学习方法..................................................................343.4本章小结.........................................................................................................37第四章图像多目标区域提取..................................................................394.1引言.................................................................................................................394.2候选目标区域的提取.....................................................................................404.2.1基于改进的光谱残差法的候选目标区域提取...................................404.2.2基于边缘检测的候选目标区域提取...................................................434.2.3基于滑动窗口的候选目标区域提取...................................................444.3候选目标区域的筛选与评价.........................................................................454.3.1利用颜色对比度筛选目标区域...........................................................45厦门大学博硕士论文摘要库4.3.2利用超像素块评价目标区域...............................................................464.4图像多目标区域提取实验.............................................................................484.5图像多目标区域提取框架.............................................................................494.6本章小结.........................................................................................................50第五章第五章第五章总结与展望总结与展望..................................................................................535.1本文总结.........................................................................................................53VI 5.2研究展望.........................................................................................................54参参参考考考文文文献献献........................................................................................55硕士在读期间科研成果介绍....................................................................63发表论文................................................................................................................63参与项目................................................................................................................63致谢............................................................................................................65附录ACorel-1k数据库介绍...................................................................67厦门大学博硕士论文摘要库VII 厦门大学博硕士论文摘要库 ContentsAbstractinChinese....................................................................................IAbstract....................................................................................................IIIContentsinChinese...................................................................................VContents....................................................................................................IXChapter1Introduction..............................................................................11.1ResearchBackgroundandsignificanceofImageRetrieval.........................11.1.1ResearchBackground...............................................................................11.1.2ResearchSignificance..............................................................................21.2ResearchStatusandHotspotsofImageRetrieval........................................21.2.1ResearchStatusofImageRetrieval..........................................................21.2.2ResearchHotspotsofImageRetrieval.....................................................31.3FormerWorkSummaryoftheResearchTeam.............................................61.4MainContentsandOrganizationalStructure...............................................7Chapter2Low-levelFeaturesextractionandcontrast..........................92.1Prologue.............................................................................................................92.2ColorFeatures...................................................................................................92.2.1ColorSpace..............................................................................................92.2.2ColorMoments.......................................................................................10厦门大学博硕士论文摘要库2.2.3ColorHistogram.....................................................................................132.3TextureFeatures.............................................................................................152.3.1DefinitionofTexture..............................................................................152.3.2TextureFeatureExtractionMethods......................................................152.3.3ClassificationExperimentsUsingVariousTextureFeatures..................212.4ShapeFeature.................................................................................................222.4.1ShapeFeatureExtractionMethods.........................................................22IX 基于目标区域的图像检索技术研究2.4.2ClassificationExperimentsUsingVariousShapeFeatures....................252.5Pseudo-semanticInformationFeature.........................................................252.5.1DefinitionofPseudo-semanticInformationFeature..............................252.5.2Pseudo-semanticInformationFeatureExtractionMethods...................262.5.3ClassificationExperimentsUsingVariousPseudo-semanticInformationFeatures............................................................................................................272.6Conclusion.......................................................................................................27Chapter3ImageFeatureFusion............................................................293.1Prologue...........................................................................................................293.2FeatureFusionofFeature-level.....................................................................293.2.1FeatureFusionBasedonSplicing........................................................293.2.2FeatureFusionBasedonSalientRegions..............................................323.3FeatureFusionofDecision-level...................................................................343.3.1SVMMultipleKernelLearningMethods..............................................343.4Conclusion.......................................................................................................37Chapter4Multi-objectRegionsExtraction..........................................394.1Prologue...........................................................................................................394.2ExtractionofCandidateObjectRegions......................................................404.2.1ExtractionofCandidateObjectRegionsBasedonImprovedSpectralResidual...........................................................................................................404.2.2ExtractionofCandidateObjectRegionsBasedonEdgeDetection.......434.2.3ExtractionofCandidateObjectRegionsBasedonSlidingWindow.....44厦门大学博硕士论文摘要库4.3ScreeningandEvaluationofCandidateObjectRegions............................454.3.1ScreeningCandidateObjectRegionsusingcolorcontrastmethod........454.3.2EvaluatingCandidateObjectRegionsusingSuperPixelsmethod........464.4Multi-objectRegionsExtractionExperiments............................................484.5Multi-objectRegionsExtractionFramework..............................................494.6Conclusion.......................................................................................................50X Degreepapersareinthe“XiamenUniversityElectronicThesesandDissertationsDatabase”.Fulltextsareavailableinthefollowingways:1.IfyourlibraryisaCALISmemberlibraries,pleaselogonhttp://etd.calis.edu.cn/andsubmitrequestsonline,orconsulttheinterlibraryloandepartmentinyourlibrary.2.Forusersofnon-CALISmemberlibraries,pleasemailtoetd@xmu.edu.cnfordeliverydetails.厦门大学博硕士论文摘要库