基于多分辨共生矩阵的纹理图像分类【荐】.pdfVIP

计算机研究与发展 ISSN1000—1239／CN1i-1777／TP JournalofComputerResearchandDevelopment 48(11)：1991—1999，2011 基于多分辨共生矩阵的纹理图像分类 钟 桦 杨晓鸣 焦李成 (智能感知与图像理解教育部重点实验室(西安 电子科技大学) 西安 710071) (hzhong@mail．xidian．edu．cn) TextureClassificationBasedonM ultires0lutiOnCo-occurrenceMatrix ZhongHua，YangXiaoming，andJiaoLicheng (KeyLaboratoryofIntelligentPerceptionandImageUnderstanding (XidianUniversity)，MinistryofEducation，Xian 710071) Abstract GraylevelCO—occurrencematrix (GLCM )iSwidely used forthedescription ofdifferent texturesbecause of its advantage ofrepresenting the texture structure． In order to hold the multiresolutionproperty andspectrum information simultaneously，GLCM isoftencomputedfrom eachwaveletsubhand，which，however，leadstomuchhigherdimensionoffeature．Toovercomethis problem ，anew featureextractingalgorithm isproposednamedmultiresolutionCO—occurrencematrix (MCM)，whosefeaturevectorhasacomparativelOW dimension．Forthetaskofdimensionreduction， thoughtheMC：M iScomputedfrom eachsubbandofundecimatedwavelettransformationasGLCM does，theparametersfortheMCM arewellchosentakingintoconsiderationthescaleandorientation ofwaveletsubbands．Inaddition，thefeaturedimensionoftheMCM canbefurtherreducedthrougha new featureselectionmethodbasedonthecorrelationanalysisofinter—subbandsandintra—subbands． PerformanceanalysisismadeindetailforthestatisticsoftheMCM andwaveletenergyfeature．The proposedMCM ismeasuredthroughtheclassificationtestontheBrodatzalbum．Experimentalresults demonstratethatMCM statisticsoutperformsothermethodssuchaswaveletenergy，GLCM ，andthe catenationofthesetwofeatures．Extensiveexperimentsonfeatureselectionshow theeffectivenessof theproposedmethod，whichcanreducethedimensionsuccessfullywithoutanylossofclassification performance． Key words texture classification； multiresolution； CO—occurrence matrix； undecimated wavelet transformatio