一种基于奇异值分解的图像匹配算法

《一种基于奇异值分解的图像匹配算法》由会员上传分享，免费在线阅读，更多相关内容在行业资料-天天文库。

1、万方数据计算机研究与发展ISSN1000—12391CN1卜1777／TPJournalofComputerResearchandDevelopment47(1)：23—32，2010一种基于奇异值分解的图像匹配算法赵峰1’2黄庆明1’2高文31(中国科学院计算技术研究所北京100190)2(中国科学院研究生院北京100049)3(北京大学信息科学与技术学院北京100871)(fzhao@jdl．ac．cn)AnImageMatchingAlgorithmBasedonSingularValueDecompositionZhaoFen91’-。HuangQingmin91一，andG

2、aoWen31(InstituteofComputingTechnology，ChineseAcademyofSciences，Beijing100190)2(GraduateUniversityofChineseAcademyofSciences，Beijing100049)3(SchoolofElectronicsEngineeringandComputerScience，PekingUniversity，Beijing100871)AbstractImagematchingplaysanimportantroleinmanyfields；suchascomputervision

3、，remotesensingandmedicalimageanalysis．Traditional—correlation—basedmatchingmethodsrequireheavycomputationtimeandtheyaresensitivetoimagerotation．Inthispaper，anewimagematchingalgorithmbasedonsingularvaluedecompositionisproposedtoefficientlymatchimagepairswitharbitraryrotationangle．Cornerpointswit

4、hdominantorientationarefirstlyextractedasfeaturepointsinthetWOimages．Theinitialsetoffeaturepointmatchesisthenobtainedbysingularvaluedecompositionofacorrespondencestrengthmatrix．Anewexpressionofthecorrespondencestrengthmatrixisintroducedtohandlemorecomplicatedimagingconditions．Eachelementofthism

5、atrix：ScomputedfromthesimilaritymeasurebetweentWOfeaturepointsbasedonnormalizedcross-correlation．Thesimilaritymeasureisinvarianttoimagerotationbytakingintoaccountthedominantorientationofthetwofeaturepoints．Finallytheepipolargeometryconstraintisimposedontheinitialmatchesinordertorejectthefalsema

6、tches．Experimentalresultsoncomplicatedrealimagesshowtheeffectivenessandrobustnessofthealgorithm．Moreover，itiseasytocombinetheproposedalgorithmwithothermatchingtechniques．ExperimentalresultswithSVD-SIFTprovetheexpansibilityandpracticabilityofthealgorithm．Keywordsimagematching；singularvaluedecomp

7、osition；normalizedcross—correlation；rotationinvariance；similaritymeasure；SVD-SIFT摘要图像匹配技术在计算机视觉、遥感和医学图像分析等领域有着广泛的应用背景．针对传统的相关匹配算法计算量大、对图像旋转敏感等问题，提出一种新的基于奇异值分解的图像匹配算法．首先在待匹配图像中分别提取带主方向的角点作为特征点，通过计算特征点间经旋转补偿的归一化互相关值建立特征点相似度矩阵，然后利用奇异