文档下载 文档下载导航
设为首页 | 加入收藏
搜索 请输入内容:  
 导航当前位置: 文档下载 > 所有分类 > IT/计算机 > 互联网 > Deep Belief Networks for Image Denoising
侵权投诉

Deep Belief Networks for Image Denoising

where isalearningrate.But,exactmaximumlikelihoodlearninginthismodelisintractablebecauseexactcomputationoftheexpectationmodelisveryexpensive.Hence,inpractice,learningisdonebyfollowinganapproximationtothegradientofadifferentobjectivefunction,calledtheContrastiveDivergence(CD)[11].

2.2DeepBeliefNetworks

Oneofthemainproblemsintrainingdeepnetworksishowtoinitializeweights.Itisdif culttooptimizetheweightsinnonlinearDeepNetworkswithmultiplehiddenlayers.Withgoodinitialweights,gradientdescentworkswell,but ndingsuchinitialweightsrequiresaverydifferenttypeofalgorithmthatlearnsonelayeroffeaturesatatime.Hintonetal.[1]introducedanewalgorithmtosolvetheaboveproblembasedonthetrainingofasequenceofRBMs.ToconstructaDBNwetrainsequentiallyasmanyRBMsasthenumberofhiddenlayersintheDBN,i.e.foraDBN

Deep Belief Networks for Image Denoising

withhhiddenlayerswehavetotrainhRBMs.TheseRBMsareplacedoneontopoftheother.Figure2givesanoverviewofthebasicconcept.Forthe rstRBM,whichconsistsoftheDBNsinputlayerandthe rsthiddenlayer,theinputofRBMisthetrainingset.ForthesecondRBM,whichconsistsoftheDBNs rstandsecondhiddenlayers,theinputistheoutputofthepreviousRBMandsoforotherRBMs.

Figure2:Left:GreedylearningastackofRBMsinwhichthesamplesfromthelower-levelRBMareusedasthedatafortrainingthenextRBM.Right:ThecorrespondingDeepBeliefNetwork

[13].

Afterperformingthislayer-wisealgorithmwehaveobtainedagoodinitializationforthehiddenweightsandbiasesoftheDBNandthentheDBNis ne-tunedwithrespecttoatypicalsupervisedcriterion(suchasmeansquareerrororcross-entropy).

3Learning

TotrainaDBNforimagedenoising,thenormalizedvaluesofanimagepixelsareused.Usingmin-maxnormalization,grayscalevalueofapixel(anintegerbetween0and255)istransformedtoa oatingpointnumberbetween0and1.Unlike rstandlastlayerofDBN,otherlayershavebinarynodes.ThemainideaistotrainaDBNsuchthatitlearnstomapnoisyimagestoimageswithlowernoiseorevenwithoutnoise.Theideacanbeimplementedbylearningthebehaviorofnoiseandimagecontentsandpresentingthesebehaviorsinsomenodesatthelastlayerofthenetwork.Thenetworkistrainedwithacollectionconsistedofbothnoisyandnoiselessimages.Weusedacriterioncalledrelativeactivitytodetectnoisenodes.Relativeactivityofeachnodeisde nedasthedifferencebetweentwovaluesofaparticularnoderesultedfromfeedingthenetworkusinganoiselessimageanditscorrespondingnoisyimage(Twoimageswithsamecontentsbutoneofthem

上一页第3页

你可能喜欢

热门文档

相关文档

站点地图 | 文档上传 | 侵权投诉 | 手机版
本站所有资源均来自互联网,本站只负责收集和整理,均不承担任何法律责任,如有侵权等其它行为请联系我们.
文档下载 Copyright 2013 doc.guandang.net All Rights Reserved.  -  email