Semantic Image Inpainting with Perceptual and Contextual Losses

《Semantic Image Inpainting with Perceptual and Contextual Losses》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、SemanticImageInpaintingwithPerceptualandContextualLossesRaymondYehChenChenTeckYianLim,MarkHasegawa-JohnsonMinhN.DoDept.ofElectricalandComputerEngineeringUniversityofIllinoisatUrbana-Champaign{yeh17,cchen156,tlim11,jhasegaw,minhdo}@illinois.eduAbstractInthispaper,wep

2、roposeanovelmethodforimageinpaintingbasedonaDeepConvolutionalGenerativeAdversarialNetwork(DCGAN).Wedefinealossfunc-tionconsistingoftwoparts:(1)acontextuallossthatpreservessimilaritybetweentheinputcorruptedimageandtherecoveredimage,and(2)aperceptuallossthatensuresaperce

3、ptuallyrealisticoutputimage.Givenacorruptedimagewithmiss-ingvalues,weuseback-propagationonthislosstomapthecorruptedimagetoasmallerlatentspace.Themappedvectoristhenpassedthroughthegenerativemodeltopredictthemissingcontent.TheproposedframeworkisevaluatedontheCelebAandSV

4、HNdatasetsfortwochallenginginpaintingtaskswithrandom80%corruptionandlargeblockycorruption.Experimentsshowthatourmethodcansuccessfullypredictsemanticinformationinthemissingregionandachievepixel-levelphotorealism,whichisimpossiblebyalmostallexistingmethods.1Introduction

5、Thegoalofinpaintingistoreconstructthemissingordamagedportionsofanimage.Ithasnu-merousapplicationssuchasrestorationofdamagedpaintingsorimageediting[3].Inpaintingforgeneralimagesisachallengingproblemasitisill-posed.Sufficientpriorinformationisrequiredinthereconstructiont

6、oachieveameaningfulandvisuallybelievableresult.Existingmethodsareoftenbasedoneitherlocalornon-localinformationtorecovertheimage.Localmethodsrelyonthepriorinformationthatexistintheinputimage.Forexample,theholesintextureimagescanbefilledbyfindingthenearestpatchesfromthesa

7、meimage[6].TotalvariationarXiv:1607.07539v2[cs.CV]14Nov2016approachestakeintoaccountofthesmoothnessofanaturalimage,whichenablessmallholesandspuriousnoisestoberemoved[22].Certainsubsetsofimagescouldalsocontainspecialproperties,suchasbeingplanar[12]orhavingalowrankstruc

8、ture[11];reconstructionresultscanbegreatlyimprovedbytakingintoconsiderationthesepriorknowledge.Theselocalmethodscanbeefficien