颜色共生矩阵的Fisher信息度量及识别应用.pdfVIP

ComputerEngineeringandApplications计算机工程与应用 20l5，51(5) 19 颜色共生矩阵的Fisher信息度量及识别应用 张 明 1,2刘传才 ，邹 健 ，卢桂馥 ZHANG Yue一，LIU Chuancai’，ZOU Jian，LU Guifu 1．南京理工大学 计算机学院，南京 210094 2．安徽工程大学 应用数理学院，安徽 芜湖 241000 3．安徽工程大学 计算机信息学院，安徽 芜湖 241000 1．SchoolofComputerScienceandTechnology,NanjingUniversityofScienceandTechnology，Nanjing210094，China 2．SchoolofM athematicsandPhysics，AnhuiPolytechnicUniversity,W uhu，Anhui241000，China 3．SchoolofComputerandInformation，AnhuiPolytechnicUniversity,W uhu，Anhui241000，China ZHANG Yue，LIU Chuancai，ZOU Jian，eta1．Fisherinformationm etricforcolorco-occurrencematricesandits recognitionapplications．ComputerEngineeringandApplications，2015，51(5)：19—22． Abstract：BasedontheRiemanniangeometryofmultinomialmanifold，anovelmethoddesignedintherfameworkofmatrix manifoldispresentedtoquantizetheinformationdivergencesbetweenColorCo—occurrenceMatrices(CCMs)forobject recognition．W iththegivencolorquantization levelsandtheneighbourhoodmodeofperpixel，themethodmodelsthe CO—occurringcolorsinanytwocolorchannelsofacolorimageasaprobabilisticrealizationofanunderlyingmultinomial distribution．BythecompactifiedCO—occurrencerfequency-basedembedding，each imageisidentifiedwithapointofa productmatrixmanifoldwhereeachfactormanifoldisendowedwiththeFisherinformationdistancemetricthatisinduced from correspondingmultinomialmanifold．Forarecognitiontask，thenearestneighbormatchingbetweentestingsamples andtrainingsamplesisfirstconductedbylabelpredictiononeveryfactormanifold，andthenfollowedbymajorityvoting on productmanifold．Theeffectivenessoftheproposedmethod isvalidatedbythepromisingrecognition resultson the GTfacedatabaseandtheCOIL一100objectdatabase． Keywords：ColorCo—occurrenceMatrices(CCMs)：multinomialdistribution；colorCO—occurrencematrixmanifold；Fisher jnf0rlTlatjondjstance 摘 要：基于多项流形的黎曼几何，提 出一个在矩阵流形框架下度量颜色共生矩阵信息差异并将其应用于 目标识别 的新方法。对于给定的颜色量化水平和每个像素局部邻域，该方法将一幅彩色图像的任意两个颜色通道中共生的 颜色建模为一个潜在的多项分布的概率实现。通过基于紧化的共生频率嵌入，可将每幅图像等同为一个积矩阵