资源描述：
《一种彩色纹理图像的分割方法》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、第24卷第9期计算机学报Vol.24No.92001年9月CHINESEJ1COMPUTERSSept.2001一种彩色纹理图像的分割方法李厚强刘政凯詹曙(中国科学技术大学电子工程与信息科学系合肥230027)摘要纹理分析一直是图像理解和计算机视觉等领域研究的重点和难点.现有的纹理分割方法大多集中于研究灰度纹理图像,文中提出一种基于分形理论和BP神经网络的彩色纹理图像分割方法.该方法将彩色图像由RGB色彩空间转换为HSI色彩空间,根据亮度计算分数维、多重分形广义维数谱q2D(q)和“空隙”等纹理特征,同时加入归一
2、化的色度和饱和度作为另外两个分类特征,采用经过有监督训练的BP神经网络作为分类器.通过对纹理图像的分割实验,结果证实该方法行之有效.关键词彩色纹理图像,图像分割,分形,神经网络中图法分类号:TP391ASegmentationMethodofColorTextureImageLIHou2QiangLIUZheng2KaiZHANShu(DepartmentofElectricalEngineering&InformationScience,UniversityofScience&TechnologyofChina
3、,Hefei230027)AbstractTextureanalysishaslongbeenanimportantresearchareaofimagecomprehensionandcomputervision.Whiletheabilityofhumantodistinguishdifferenttexturesisapparent,theautomateddescription,recognition,andsegmentationofthesesamepatternshaveproventobequit
4、ecomplex.Mostexistingtexturesegmentationapproachesfocusongraytextureimages,whilethispaperpresentsacolortextureimagesegmentationmethod,whichisbasedonfractaltheoryandBPneuralnetwork.Therearetwoimportantaspectsoftextureimagesegmentationprocedures,namelyfeatureex
5、tractionandsegmentation.Mosttexturefeaturesoftenbeingusedinsegmentationarebasedonsuchmodelsasstructuralmodel,statisticmodel,filtermodel,ran2domfieldmodelandfractalmodel.Thetexturefeaturesstudiedinthispaperareallbasedonthefractaltheory.Thischoiceismotivatedbyt
6、heobservationthatthefractaldimension(FD)isrel2ativelyinsensitivetoanimagescaling,andshowsastrongcorrelationwithhumanjudgementofsurfaceroughness.Butfractaldimensionalonedoesnotprovidesufficientinformationtodescribeandsegmentnaturaltextures.Fractalsetsmaysharet
7、hesamefractaldimensionandyethavestrikinglydifferentappearancesortextures.Thusmultifractaltheoryisintroducedinthepaper.ThetexturefeaturebasedonmultifractaltheoryisspectrumfunctionD(q),whichcancharacter2izeefficientlynaturaltextureseveniftheyarequitesimilar.Ane
8、wclassoftexturemeasuresbasedontheconceptoflacunarityisdefined.Thesecond2orderstatisticlacunarityissmallwhenthetextureisdense,andlargewhenthetextureiscoarse.Basedontheabovework,thispaperpr
