基于稀疏编码的视觉模型及其应用

《基于稀疏编码的视觉模型及其应用》由会员上传分享，免费在线阅读，更多相关内容在学术论文-天天文库。

1、上海交通大学硕士学位论文基于稀疏编码的视觉模型及其应用姓名：周期申请学位级别：硕士专业：计算机软件与理论指导教师：张丽清20090201ii组合等等较好的特性，因此成为了研究的热点。本文在基于已有的非负稀疏编码的基础上，加入了拓扑先验。同时，根据非负矩阵分解的乘性迭代公式的推导，推导出对于拓扑先验的乘性迭代公式。使用乘性迭代公式，算法的效率会更高，同时实验结果表明，该模型同样能够得到超完备，抗噪声的拓扑的结构，更好的解释的生物现象。关键词：稀疏编码，计算机视觉，超完备表征，视觉初级皮层ABSTRACTVisionperceptio

2、nisoneoftheintriguingproblemswhichattractshumanbeingsforalongtime.Itisachallengetodesignandconstructafastande±cientvisualsystemthatcancompetewithhumanvisualsystems.Withthedevelopmentofthescienceandtechnology,includingtherapidimprovementofcomputerability,thepopularizat

3、ionofhighprecisionandspeedimageacquisitionhardware,emergenceofnewvisionalgorithmandmoredatafromneurobiologyexperiments,designingasystemthatcanbecompetentforhumanvisionbasictaskssuchasfactdetectionandrecognitionisstillachallengeproblem.Theapproachtolearnandsimulatethef

4、eaturesandinternalworkingmechanismofbiologicalvisionsystemsistreatedasae®ectivewaytosolvetheproblem.Inthisthesis,webrie°yintroducethehumanvisionsystem-especiallytheprimatevisualcortex-anditsstructureandcharacteristics.Meanwhile,wedescribethemodernmodelingtools,includi

5、ngprobabilisticmodelandlineargen-erativeframework.Basedonthesetools,wealsodepicttheprimarycomponentanalysis,independentcomponentanalysis,sparsecodingandtheirmodelingonhumanprimatevisualsystems.Thefeatureslearnedbythesemethodscanre-vealsomeimportantcharacteristicssucha

6、ssparse,overcompleteandtopography.Themaincontributionsofthepaperare:1.SparsitypriorcriterionThesparsepropertyisconsideredasoneoftherea-sonsaccountingforthee±ciencyandhighspeedofhumanbrain.However,howtode¯nesparsityisstillnotwellde¯ned.Therearestillsomesubjectfactsinju

7、dgingsomedistributionsaresparewhilesomeothersarenot.Thepaperderivesthatsomepriorsinclassicalsparsecodingmodelwilltendtospreadaneuron'sactivationstotherestofneurons'toreducetheobjec-tivevalue.Thisphenomenonbreakstheassumptionofsparse.Wede¯nethevaluereducedas"duplicateb

8、onus"anddeducethatthesparsefunc-tionshouldsatisfysubadditiveproperty.Weevaluatedi®erentdistributionunderthiscriterionandvali