cvpr18-Separating Style and Content for Generalized Style Transfer 外文学习材料

《cvpr18-Separating Style and Content for Generalized Style Transfer 外文学习材料》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、SeparatingStyleandContentforGeneralizedStyleTransferYexunZhangYaZhangWenbinCaiShanghaiJiaoTongUniversityShanghaiJiaoTongUniversityMicrosoftzhyxun@sjtu.edu.cnyazhang@sjtu.edu.cnwenbca@microsoft.comAbstractStyleStyleReferenceSetRepresentation…Style��EncoderOutputNeuralstyletransferhas

2、drawnbroadattentioninre-��W��WDecodercentyears.However,mostexistingmethodsaimtoexplic-…Content��itlymodelthetransformationbetweendifferentstyles,andEncoderMixerContentReferenceSetContentthelearnedmodelisthusnotgeneralizabletonewstyles.RepresentationWehereattempttoseparatetherepresent

3、ationsforstylesFigure1.TheframeworkoftheproposedEMDmodel.andcontents,andproposeageneralizedstyletransfernet-cently,severalvariationsofgenerativeadversarialnetworksworkconsistingofstyleencoder,contentencoder,mixerand(GANs)[14,28]areintroducedbyaddingadiscriminatordecoder.Thestyleencod

4、erandcontentencoderareusedtothestyletransfernetworkwhichincorporatesadversarialtoextractthestyleandcontentfactorsfromthestyleref-losswithtransferlosstogeneratebetterimages.However,erenceimagesandcontentreferenceimages,respectively.thesestudiesaimtoexplicitlylearnthetransformationfrom

5、Themixeremploysabilinearmodeltointegratetheaboveacertainsourcestyletoagiventargetstyle,andthelearnedtwofactorsandfinallyfeedsitintoadecodertogeneratemodelisthusnotgeneralizabletonewstyles,i.e.retrainingimageswithtargetstyleandcontent.Toseparatethestyleisneededfortransformationsofnewst

6、yleswhichistime-featuresandcontentfeatures,weleveragetheconditionalconsuming.dependenceofstylesandcontentsgivenanimage.DuringInthispaper,weproposeanovelgeneralizedstyletrans-training,theencodernetworklearnstoextractstylesandfernetworkwhichcanextendwelltonewstylesorcontents.contentsfr

7、omtwosetsofreferenceimagesinlimitedsize,Differentfromexistingsupervisedstyletransfermethods,onewithsharedstyleandtheotherwithsharedcontent.whereanindividualtransfernetworkisbuiltforeachpairThislearningframeworkallowssimultaneousstyletrans-ofstyletransfer,theproposednetworkrepresentse

8、achstylefera