Pixel Recursive Super Resolution

Pixel Recursive Super Resolution

ID:39715943

大小:4.94 MB

页数:21页

时间:2019-07-09

Pixel Recursive Super Resolution_第1页
Pixel Recursive Super Resolution_第2页
Pixel Recursive Super Resolution_第3页
Pixel Recursive Super Resolution_第4页
Pixel Recursive Super Resolution_第5页
资源描述:

《Pixel Recursive Super Resolution》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、PixelRecursiveSuperResolutionRyanDahlMohammadNorouziJonathonShlensGoogleBrainfrld,mnorouzi,shlensg@google.comAbstract88input3232samplesgroundtruthWepresentapixelrecursivesuperresolutionmodelthatsynthesizesrealisticdetailsintoimageswhileenhancingth

2、eirresolution.Alowresolutionimagemaycorrespondtomultipleplausiblehighresolutionimages,thusmodelingthesuperresolutionprocesswithapixelindependentcon-ditionalmodeloftenresultsinaveragingdifferentdetails–henceblurryedges.Bycontrast,ourmodelisabletorepre

3、-sentamultimodalconditionaldistributionbyproperlymod-elingthestatisticaldependenciesamongthehighresolutionimagepixels,conditionedonalowresolutioninput.WeemployaPixelCNNarchitecturetodefineastrongpriorovernaturalimagesandjointlyoptimizethispriorwithade

4、epconditioningconvolutionalnetwork.Humanevalua-tionsindicatethatsamplesfromourproposedmodellookmorephotorealisticthanastrongL2regressionbaseline.1.IntroductionTheproblemofsuperresolutionentailsartificiallyen-largingalowresolutionphotographtorecoverapl

5、ausi-Figure1:Illustrationofourprobabilisticpixelrecursiveblehighresolutionversionofit.Whenthezoomfactorsuperresolutionmodeltrainedend-to-endonadatasetofarXiv:1702.00783v1[cs.CV]2Feb2017islarge,theinputimagedoesnotcontainalloftheinfor-celebrityfaces.T

6、heleftcolumnshows88lowresolutionmationnecessarytoaccuratelyconstructahighresolutioninputsfromthetestset.Themiddleandlastcolumnsshowimage.Thus,theproblemisunderspecifiedandmanyplau-3232imagesaspredictedbyourmodelvs.thegroundsiblehighresolutionimagese

7、xistthatmatchthelowresolu-truth.Ourmodelincorporatesstrongfacepriorstosynthe-tioninputimage.Thisproblemissignificantforimprovingsizerealistichairandskindetails.thestate-of-the-artinsuperresolution,andmoregenerallyforbuildingbetterconditionalgenerative

8、modelsofimages.Asuperresolutionmodelmustaccountforthecomplexnewimagedetailsthatappearplausibletoahumanob-variationsofobjects,viewpoints,illumination,andocclu-server.Generatingrealistichighresolutionimagesisnotsions,especiallyasthezoomfactorincreases.

当前文档最多预览五页,下载文档查看全文

此文档下载收益归作者所有

当前文档最多预览五页,下载文档查看全文
温馨提示:
1. 部分包含数学公式或PPT动画的文件,查看预览时可能会显示错乱或异常,文件下载后无此问题,请放心下载。
2. 本文档由用户上传,版权归属用户,天天文库负责整理代发布。如果您对本文档版权有争议请及时联系客服。
3. 下载前请仔细阅读文档内容,确认文档内容符合您的需求后进行下载,若出现内容与标题不符可向本站投诉处理。
4. 下载文档时可能由于网络波动等原因无法下载或下载错误,付费完成后未能成功下载的用户请联系客服处理。