eatureextractionalgorithmsexaminedinpreviousstudies..MethodologyInthenewapproachweproposed,palmprintisidentifiedbythetotalenergylevelinsectorsofellipticalhalf-ringsofthepalmimage.Theproposedalgorithmallowstheextractedpalmregiontovaryproportionallywithpalmsize.Atthesametime,itcancopewithslightvariationsintheextractedregionwithinaperson,i.e.evenwhentwofeaturevectorswereextractedfromdifferentpalmsizelsofourproposedmethod.SectiondiscussestheexperimentalsetupwiththeresultsshowninSection.Sectionconcludesthemajorfindingsandcontributionsofthispaper..RelatedWorkThestudiesofpalmprintswerefirstcarriedoutoninkedpalmprintimages[].Notonlythattheimagecollectionprocessistediousandunrealistic,thehollowedcentralpartofthepalmisoftenmissing.Withtheadvanceoftechnology,inklesspalmprintimagescannowbecaptured,byeitherscanningtechnologiesorCCDcameras.PalmimagescapturedbyCCDcameras,whichisthetechnologyusedinthisstudy,areoflowerresolutionbutinturnrequiredlessprocessingtime.Apartfromutilizingthestructuralpropertiesofpalmprint[][],othertypicalfeaturesthathavebeenstudiedinclude:fuzzydirectionalelementenergyfeatures,adoptedfromChinesecharactersrecognitionalgorithms[]；featuresinFourierspace[]；eigenfeaturesretainedafterperformingdimensionalityreductionbyKarhunen-Loevetransform[]orFisher’slineardiscriminant[]；and,statisticalfeaturesobtainedbytextureanalysisusingGaborfilters[]orwaveletstransform[][].Inallthesepreviousstudies,subjectsarerequiredtoplacetheirhandsonacontactsurface.Someevenfixedsubjects’handposturebypegs[].Featuresareonlyextractedfromafixedsizesquareblockofthecentralpartofthepalm.Sincepalmsizevariesgreatlyamongstpeople,usingonlyafixedsizepalmregiononallpalmprintsactuallyneglectsalotofdistinctiveanduniqueinformationontheotherpartsofthepalmprint.Thisshortfallissignificantlyrelevantwhenthesizeofthepalmprintislarge(seeFigure).Wethereforeproposedanewmethodindividingandarrangingthefeaturesextractedfrompalmprintsuchthattheutilizationregionofthepalmprintisgreatlyincreasedandmorefeaturescanbeextracted.Awavelet-basedtechniqueischosentoanalyzepalmprints,i.e.bythesimpleHaarwavelettransform,whichhasbeendemonstratedtoproducethebestresultsamongstvariouswavelettransforms[].Itisnotedthattheproposedmethodcanbeeasilyintegratedintootherfeatureextractionalgorithmsexaminedinpreviousstudies..MethodologyInthenewapproachweproposed,palmprintisidentifiedbythetotalenergylevelinsectorsofellipticalhalf-ringsofthepalmimage.Theproposedalgorithmallowstheextractedpalmregiontovaryproportionallywithpalmsize.Atthesametime,itcancopewithslightvariationsintheextractedregionwithinaperson,i.e.evenwhentwofeaturevectorswereextractedfromdifferentpalmsize本图中应用，其结果就是从图像中最小限度减去一个维物体投射到维图像中非均匀光照影响。其次，由于高斯窗矩形波对噪声敏感特性，高斯窗被用来光滑哈尔小波图像。一个一级图像分解是通过哈尔小波来完成。通过图像包括水平，垂直，对角这三个细节，我们了解到，一个光滑工具被用来消除噪声。研究发现，大部分低频率成分来源于皮下红肿，并且为了更好识别，应该从特征中排除这部分。因此，像素与频率值之间标准差应该为零。像素其他值被投射到一个对数式中，以至于可以减小两幅图像中频率成分大小绝对差异。那就是：yxIstdyxIifwoyxIstdyxIjijijiyxI,,,..,,,ln,这里I（xi,yi）是细节图像频率值。最后，一个环形形式中乘操作工具被用来加强检测手掌连通性和厚度。图中显示了这个过程。图.掌纹识别哈尔波变换.特征向量组成每一个细节图像分成了非重叠椭圆环形。这些椭圆都以同一个点为中心，随着面积增加倍，椭圆主轴心和次轴心会以其直接内心为基准扩大两倍。每一个环形都会被分成不同数量部分。最深处椭圆分成了个部分，然而要从这里移动出去，每一个外环将比它内层多两部分，如图（a）所示。每一环节能量水平就是频率成分绝对总和除以每一环节中像素数量，并且用它来构造特征向量。特征能量排列就是对于所有三层细节图形而言，内层能量水平高于外层能量水平。如图（b）所示，这个排列确保了当两个不同程度特征向量进行相互比较时候，它们之间逐点比较就是两个不同手掌图像中相同特殊区域之间特性比较。图.特征向量.匹配分值计算不考虑原始图像大小，我们把处理过程中手掌图形分成了宽度相同椭圆半环，然而不同大小手掌面积将导致不同长度特征矢量。由于在手掌拉伸过程中有出现变化可能性，所以合成最大手掌面积可能会在同一物体中有所变化。因此，距离度量使用必须可以公平地通过不相等维度来比较两个特征向量。分值计算作为两个特征向量之间绝对差异平均值。如果Vi代表Ni元素一个特征向量，那么两个图像之间分值计算如下：jiNNnjiNNnfeatureVnfeatureVjiScoreji,min||,,min、实验装置我们已经从个年龄在到岁范围内群体中得到了他们手掌图像。大约是女性，是中国人。手掌图像是以x以像素为单位分辨率和位颜色而获取。人中每十幅图像左手和右手是从一个有个右手围绕垂直轴翻转并且储存为另一个物体数据库中被提取出来。我们对掌纹识别研究不同于以往研究，在以往掌纹识别研究中，我们要求受试者把手放在一个可以接触表面上，并且用钉子来固定手姿势，而我们现在研究是使用了一个在工作时不受限制非接触式捕捉系统。我们仅仅要求受试者把手放在一个酥软平面上，手掌向上并且五指分开。一个在桌子上保持固定距离CCD相机用来捕获完整手掌图像。这个装置避免了为了保持在接触器件中获得高质量捕获图片而频繁清洗接触表面麻烦，并且给使用者提供了一个方便和舒适环境。通过数据库使用，实验识别和验证被执行。为了可以识别，十倍和二倍交叉验证方法和一阶临近取样分类器同时使用。为了验证，每张图片都要和其余图片相比较，这就导致了真正分数和假冒分数。该算法性能将会和以往采用固定大小面积区域并且分成多个非重叠方块研究进行比较。该测试使用了奔腾IVMHz处理器和MB内存MATLAB.系列。、成果该算法使用图像处理和构建相应特征向量平均时间比以往算法慢了..秒，但是该算法中每张图像特征数量比以往少了。对于十倍一阶临近取样分类器，该方法精度可以达到.。对于二倍分类器，精度下降到.。结果表明，该方法比以往算法更好，因为以往算法中十倍分类器和二倍分类器精度分别是.和.。新方法和旧方法平均误差率分别是.和.，比较总结见表。表.性能比较、总结本文提出了一种定位分割掌纹中感兴趣区域新方法。这一新方法利用了个人可以达到特征提取最大掌纹区域。更重要是，它能应付从同一个人掌纹图像中出现旋转，转换和尺寸差异等微小变化。特征向量排列是通过对两个不同手掌中相同特殊区域进行逐点比较得到。当我们把每个物体捕获用作个物体数据库来训练时，随着精度高达到.，试验成果是有希望。甚至当仅仅只有五个图片时，精度水平仍然可以保持在.。验证可知，新方法平均误差率可以达到.，而从固定大小手掌区域中提取出非重叠方块平均误差率却达到了.。、感谢该研究由香港科技大学资助信和软件研究所提供，资助编码为SSRI/.EG。lsofourproposedmethod.SectiondisussestheexperimentalsetupwiththeresultsshowninSection.Sectionconcludesthemajorfindingsandcontributionsofthispaper..RelatedWorkThestudiesofpalmprintswerefirstcarriedoutoninkedpalmprintimages[].Notonlythattheimagecollectionprocessistediousandunrealistic,thehollowedcentralpartofthepalmisoftenmissing.Withtheadvanceoftechnology,inklesspalmprintimagescannowbecaptured,bye外文资料翻译所在学院：通信与信息工程学院专业(班级)：通信工程（）学生姓名：指导教师：年月日ANewMethodinLocatingandSegmentingPalmprintintoRegion-of-InterestC.Poon,D.C.M.Wong,H.C.ShenDept.ofComputerScienceTheHongKongUniversityofScience&TechnologyHongKong{helens@cs.ust.hk}AbstractVarioustechniquesinanalyzingpalmprinthavebeenproposedbuttothebestofourknowledge,nonehasbeenstudiedontheselectionanddivisionoftheregion-ofinterest(ROI).Previousmethodswerealwaysappliedonlytoafixedsizesquareregionchosenasthecentralpartofthepalm,whichwerethendividedintosquareblocksforextractionoflocalfeatures.Inthispaper,weproposedanewmethodinlocatingandsegmentingtheROIforpalmprintanalysis,wheretheselectedregionvarieswiththesizeofthepalm.Insteadofsquareblocks,theregionisdividedintosectorsofellipticalhalf-rings,whicharelessaffectedbymisalignmentduetorotationalerror.Moreimportantly,ourarrangementofthefeaturevectorsensuresthatonlyfeaturesextractedfromthesamespatialregionoftwoalignedpalmswillbecomparedwitheachother.Encouragingresultsobtainedfavortheuseofthismethodinthefuturedevelopmentofpalmprintanalysistechniques..IntroductionBiometrics,identificationofapersonbyhis/herphysiologicalorbehavioralcharacteristics,hasbecomeincreasinglyprevalentinmodernidentificationandverificationsystems[][].Ofallthebiometricsstudied,palmprinthasanadvantageoverotherbiometricssuchasvoiceandfacerecognitionwhereuniquenessbetweenpeopleisdoubtful[]orfingerprintandirispatternwherehigh-resolutionimagesarerequired(e.g.overdpi).Palmprintsareuniquebetweenpeopleandrelativelylowresolutionimageswillsuffice(lessthandpi)[][].However,thereisamajorshortfallinthepreviousalgorithms.Thatistheyonlyutilizeafixedareaofapalmforidentificationregardlessoftheactualpalmsize.Obviously,alotofinformationhasbeenoverlooked.Twopalmsarecapturedfromthesamedistanceandwiththesameresolution.Afixedsizeregion-of-interest(ROI)(reddottedsquare)isconsideredtobesmallonlargepalms.Figure.PalmsizedifferenceintwoindividualsInthispaper,weproposeanewmethodinpalmprintidentificationthatallowstheROIinanalyzingapalmimagetovarywiththeactualpalmsize(seeFigure,bluesolidsquare).Inthefollowingsections,relatedworkinthestudyofpalmprintispresentedinSection.Sectionprovidesthedetailsofourproposedmethod.SectiondiscussestheexperimentalsetupwiththeresultsshowninSection.Sectionconcludesthemajorfindingsandcontributionsofthispaper..RelatedWorkThestudiesofpalmprintswerefirstcarriedoutoninkedpalmprintimages[].Notonlythattheimagecollectionprocessistediousandunrealistic,thehollowedcentralpartofthepalmisoftenmissing.Withtheadvanceoftechnology,inklesspalmprintimagescannowbecaptured,byeitherscanningtechnologiesorCCDcameras.PalmimagescapturedbyCCDcameras,whichisthetechnologyusedinthisstudy,areoflowerresolutionbutinturnrequiredlessprocessingtime.Apartfromutilizingthestructuralpropertiesofpalmprint[][],othertypicalfeaturesthathavebeenstudiedinclude:fuzzydirectionalelementenergyfeatures,adoptedfromChinesecharactersrecognitionalgorithms[]；featuresinFourierspace[]；eigenfeaturesretainedafterperformingdimensionalityreductionbyKarhunen-Loevetransform[]orFisher’slineardiscriminant[]；and,statisticalfeaturesobtainedbytextureanalysisusingGaborfilters[]orwaveletstransform[][].Inallthesepreviousstudies,subjectsarerequiredtoplacetheirhandsonacontactsurface.Someevenfixedsubjects’handposturebypegs[].Featuresareonlyextractedfromafixedsizesquareblockofthecentralpartofthepalm.Sincepalmsizevariesgreatlyamongstpeople,usingonlyafixedsizepalmregiononallpalmprintsactuallyneglectsalotofdistinctiveanduniqueinformationontheotherpartsofthepalmprint.Thisshortfallissignificantlyrelevantwhenthesizeofthepalmprintislarge(seeFigure).Wethereforeproposedanewmethodindividingandarrangingthefeaturesextractedfrompalmprintsuchthattheutilizationregionofthepalmprintisgreatlyincreased        1外文资料翻译所在学院：通信与信息工程学院专业(班级)：通信工程（072）学生姓名：指导教师：2011年5月9日2ANewMethodinLocatingandSegmentingPalmprintintoRegion-of-InterestC.Poon,D.C.M.Wong,H.C.ShenDept.ofComputerScienceTheHongKongUniversityofScience&TechnologyHongKong{helens@cs.ust.hk}AbstractVarioustechniquesinanalyzingpalmprinthavebeenproposedbuttothebestofourknowledge,nonehasbeenstudiedonthes

文档预览结束

找同类相似 下载到手机