下载本文档
版权说明:本文档由用户提供并上传,收益归属内容提供方,若内容存在侵权,请进行举报或认领
文档简介
一种基于生成对抗网络的火焰图像场景迁移新方法Abstract:Withthedevelopmentofdeeplearningtechnology,imagestyletransferhasbecomeanimportantresearcharea.Flameimagescenetransferhasgreatresearchsignificanceandpracticalvalue.Inthispaperweproposeanewmethodbasedongenerativeadversarialnetwork(GAN)forflameimagescenetransfer.Thismethodconsistsofamultiscalegeneratorandadiscriminator.Thegeneratorisdesignedwithskipconnectionarchitecturetotransfertheflameimagestyletothetargetimagewhilepreservingthestructuraldetails.Thediscriminatorisusedtodistinguishthegeneratedimagefromtherealimageandguidethegeneratortoproducemorerealisticimages.Experimentsshowthatourmethodcaneffectivelytransfertheflameimagestyletothetargetimageandachieveagoodvisualeffect.nscenetransferistheprocessoftransferringtheflameimagestyletoatargetimagewhilemaintainingthestructuraldetailsofthetargetimageIthaswideapplicationsinmoviespecialeffects,videogamesandvirtualreality.WiththedevelopmentofdeeplearningtechnologyimagestyletransferhasbecomeanimportantresearchareaTraditionalmethodsofimagestyletransfermainlyrelyontexturesynthesiswhichistime-consumingandcomputationallyexpensive.Inrecentyearsgenerativeadversarialnetworks(GANs)haveemergedasapromisingmethodforimagestyletransfer.InthispaperweproposeanewmethodbasedonGANforflameimagescenetransferOurmethodisdesignedwithamulti-scalegeneratorandadiscriminatorwhichcaneffectivelytransfertheflameimagestyletothetargetimageandachieveagoodvisualeffect.agestyletransferhasbeenextensivelystudiedinrecentyearsGatysetalproposedaneuralalgorithmforimagestyletransfer,whichusestheGrammatrixoffeaturemapstorepresenttheimagestyleandoptimizethecontentandstyleseparately.However,thismethodiscomputationallyexpensiveandcannothandlelargeimages.Johnsonetalproposedafastneuralstyletransfermethod,whichusesapre-traineddeepconvolutionalneuralnetworktotransferthestyleofanimagetootherimages.However,thismethodisalsolimitedbythesizeoftheinputimageTosolvetheseproblems,severalGAN-basedmethodshavebeenproposedforimagestyletransfer.GANsconsistofageneratorandadiscriminatorThegeneratoristrainedtogeneraterealisticimages,whilethediscriminatoristrainedtodistinguishthegeneratedimagesfromtherealones.GAN-basedmethodshaveshowngreatsuccessinimagestyletransfer.Isolaetal.proposedaconditionalGANforimage-to-imagetranslation,whichcantransfervariousimagestyles,suchaschangingtheseasonofalandscapeimageortransferringthestyleofanimagefromdaytonight.InthispaperweproposeanewmethodbasedonGANforflameimagescenetransferOurmethodconsistsofamulti-scalegeneratorandadiscriminatorThegeneratorisdesignedwithskip-connectionarchitecturetotransfertheflameimagestyletothetargetimagewhilepreservingthestructuraldetails.Thediscriminatorisusedtodistinguishthegeneratedimagefromtherealimageandguidethegeneratortoproducemorerealisticimages.ThearchitectureofourgeneratorisshowninFigure1.Ourgeneratortakesthetargetimageasinputandoutputsthegeneratedimagewiththeflameimagestyle.Thegeneratorconsistsofanencoder,adecoderandskipconnectionsbetweencorrespondinglayersintheencoderanddecoderTheencoderofourgeneratorisdesignedwithamultiscalearchitecturetocapturethefeaturesatdifferentscales.Thedecoderofourgeneratorisdesignedwithadeconvolutionalarchitecturetoupsamplethefeaturesandgeneratethetargetimagewiththeflameimagestyle.ThearchitectureofourdiscriminatorisshowninFigure2.Ourdiscriminatortakesthegeneratedimageandtherealimageasinputandoutputsabinaryclassificationresultindicatingwhethertheinputimageisrealorfake.Thediscriminatorconsistsofmultipleconvolutionallayersfollowedbyafullyconnectedlayer.Theoutputofthediscriminatorisascalarvalue,whichrepresentstheprobabilityoftheinputimagebeingreal.TotrainourGANmodel,wedefinethefollowinglossfunctions:AdversariallossTheadversariallossisusedtotrainthediscriminatortodistinguishthegeneratedimagefromtherealimageandtrainthegeneratortogeneratemorerealisticimages.ThedversariallossisdefinedasLadvElogDrealElogDG(target)))]WhereDdenotesthediscriminatorandGdenotesthegenerator.ContentlossThecontentlossisusedtomeasurethesimilaritybetweenthegeneratedimageandthetargetimageintermsofcontent.Thecontentlossisdefinedas:L_content=||G(target)-target||1StylelossThestylelossisusedtomeasurethesimilaritybetweenthegeneratedimageandtheflameimageintermsofstyle.Thestylelossisdefinedas:LstyleGramGflameGramGtargetWhereGramdenotestheGrammatrixoffeaturemaps.ThetotallossofourGANmodelis:LadvLadvcontentLcontentλ_style*L_styleWhereλ_adv,λ_content,andλ_stylearehyperparameters.WeevaluateourmethodontheCIFAR-10dataset,whichcontains00trainingimagesand10,000testingimages.Werandomlyselect0testingimagesasourtargetimagesandusetheflameimagefromtheinternetasourflameimage.WetrainourGANmodelontheremainingimagesandtestitonthetestingset.TheresultsofourmethodareshowninFigure3.OurmethodcaneffectivelytransfertheflameimagestyletothetargetimageandmaintainthestructuraldetailsofthetargetimageThegeneratedimagesarevisuallypleasingandhaveagoodlevelofrealism.WecompareourmethodwiththeneuralalgorithmforimagestyletransferproposedbyGatysetal.andthefastneuralstyletransfermethodproposedbyJoh
温馨提示
- 1. 本站所有资源如无特殊说明,都需要本地电脑安装OFFICE2007和PDF阅读器。图纸软件为CAD,CAXA,PROE,UG,SolidWorks等.压缩文件请下载最新的WinRAR软件解压。
- 2. 本站的文档不包含任何第三方提供的附件图纸等,如果需要附件,请联系上传者。文件的所有权益归上传用户所有。
- 3. 本站RAR压缩包中若带图纸,网页内容里面会有图纸预览,若没有图纸预览就没有图纸。
- 4. 未经权益所有人同意不得将文件中的内容挪作商业或盈利用途。
- 5. 人人文库网仅提供信息存储空间,仅对用户上传内容的表现方式做保护处理,对用户上传分享的文档内容本身不做任何修改或编辑,并不能对任何下载内容负责。
- 6. 下载文件中如有侵权或不适当内容,请与我们联系,我们立即纠正。
- 7. 本站不保证下载资源的准确性、安全性和完整性, 同时也不承担用户因使用这些下载资源对自己和他人造成任何形式的伤害或损失。
最新文档
- 玉溪师范学院《光学》2021-2022学年第一学期期末试卷
- 2024年TFT系列偏光片合作协议书
- 广西壮族自治区桂林市第十八中2024年高三第二次适应性考试数学试题
- 2024年耐高温合成云母层压板项目发展计划
- 盐城师范学院《新能源热利用与热发电原理及系统》2023-2024学年期末试卷
- 2024公司授权委托书合同
- 浙教版五年级上册数学第一单元 小数的意义与加减法 测试卷及完整答案(必刷)
- 2024年塑料挤吹中空成型机项目合作计划书
- 2024年内螺纹球阀项目发展计划
- 2024房产赠与合同范本写
- 新教材·气象灾害之洪涝灾害(公开课)课件
- 欧盟允许使用的食品添加剂
- 部编版六年级(下)语文写人记事类阅读复习检测题(含答案)
- 人际交往能力自测量表
- (完整版)感染性疾病科设置要求
- 旅游地理课件:旅游规划及旅游线路设计
- 河北省承德市各县区乡镇行政村村庄村名居民村民委员会明细
- 灾害现场检伤分类-课件
- 日文简历模板履歴书(JIS规格)
- (完整)E级GPS控制测量技术设计书
- 疗养院建筑设计规范(含条文说明)
评论
0/150
提交评论