[PDF] [PDF] Generative Adversarial Networks for Photo - CVF Open Access

[PDF] Transformer une photo en caricature - JPPAINT

Adobe PhotoShop Effets Spéciaux Transformer une photo en caricature 1 Choisir une photo bien nette et contrastée et sélection- ner le menu 

[PDF] Réaliser une image genre bande dessinée

Dans photoshop, vous commencez par dupliquer l'image sous forme de calque : Vous allez donc dans "Image", "Réglages" et "Négatif pour transformer ce 

[PDF] Photoshop-5pdf - atelier-multimedia-brestfr

point il est simple de créer un titre de type cartoon comme ci-dessous : 6 Remarque : pour les images CMJN, l'option Mélange maximal transforme tous les

[PDF] 100 tutos Photoshop (French Edition) - livre gratuit

À tous ceux qui souhaitent parfaire leurs photos avec Photoshop Aux utilisateurs de Filtre radial et Upright (par le panneau Transformation) L'icône ci-dessus 

[PDF] Codex Illustrator version complète

de peinture comme l'eau et le gouache, l'image numérique transforme les création des images bitmaps (comme Photoshop et Windows Paint), des images  

[PDF] Photoshop - Cyber-base de Decize

Photoshop – Fiche 10 – Ajouter une bulle de BD à un personnage – Les formes personnalisées allons prendre pour exemple une simple photo à laquelle

[PDF] Photoshop CS5 - Campus Bruxelles

Dessiner un cartoon Transformer une photo en peinture avec l'outil Pinceau Photoshop est le logiciel leader depuis des années dans le monde

[PDF] Support de cours PHOTOSHOP

Photoshop permet de coloriser une image, de l'étoffer, d'y incorporer du texte, En outre, ce logiciel de traitement d'images, nous offre la possibilité de réaliser transformer une image "CMJN" en quadrichromie composée de tons directs

[PDF] Generative Adversarial Networks for Photo - CVF Open Access

photos and cartoon images for training, which is easy to use ware such as Instagram and Photoshop sired manifold transformation, and (2) the content loss

[PDF] Comment créer un GIF avec Photoshop ? - Votre Assistante

11 oct 2018 · Créer un GIF Photoshop avec des images des articles de blog, il faut tout d' abord que vous ayez la vidéo à transformer enregistrée sur

[PDF] transformer photo en dessin photoshop

[PDF] transformer photo en bd gratuit

[PDF] phylogénie moléculaire

[PDF] méthode neighbor joining

[PDF] transformer une photo en coloriage

[PDF] mécanique quantique cours pdf

[PDF] transformer photo en dessin couleur photoshop

[PDF] mécanique quantique l2 pdf

[PDF] photoshop effet dessin crayon

[PDF] mecanique quantique 2

[PDF] l'analyse des textes littéraires une méthodologie complète

[PDF] technique collage artistique

[PDF] activité collage maternelle

[PDF] tableau avec collage papier

[PDF] dessin collage papier

CartoonGAN:Generativ eAdversarialNetworksforPhotoCartoonization

YangChen

TsinghuaUniv ersity,China

CardiffUniversity, UK

Yukun.Lai@cs.cf.ac.ukYong-JinLiu

TsinghuaUniv ersity,China

liuyongjin@tsinghua.edu.cn

Abstract

Inthispaper ,wepr oposeasolutiontotransforming pho- tosofr eal-worldscenesinto cartoonstyleimages,whic his valuableandc hallengingincomputer visionandcomputer graphics.Oursolutionbelongsto learningbasedmethods, whichhaverecentlybecome populartostylize imagesin artisticformssuc haspainting .However,existing meth- odsdonot producesatisfactory resultsfor cartoonization, duetothe factthat(1) cartoonstyleshave uniquechar ac- teristicswithhigh levelsimplification andabstraction, and (2)cartoonima gestend tohaveclearedges,smoothcolor shadingandr elativelysimplete xtures,whichexhibit signif- icantchalleng esfortexture-descriptor-basedlossfunctions usedine xistingmethods.In thispaper,wepropose Car- toonGAN,ag enerativeadver sarialnetwork(GAN)frame- workforcartoon stylization.Ourmethod takesunpair ed photosandcartoon images fortraining ,whichiseasyto use.Twonovel lossessuitableforcartoonizationare pro- posed:(1)a semanticcontentloss, whichis formulatedas asparse regularizationinthehigh-le velfeaturemapsof theVGGnetwork tocopewith substantialstylevariation betweenphotosand cartoons,and(2) anedge-pr omoting adversariallossforpreserving clearedges. Wefurther in- troduceaninitializationphase, toimpro vethecon vergence ofthenetwork tothetar getmanifold. Ourmethodis also muchmoreefficient totrainthanexisting methods.Exper- imentalresults showthatourmethodisable togener ate high-qualitycartoonima gesfr omreal-worldphotos(i.e., followingspecificartists" stylesandwith clearedges and smoothshading)and outperformsstate-of-the-artmethods.

1.Introduction

Cartoonsarean artisticformwidely usedinour daily

life.Inaddition toartisticinterests, theirapplicationsrange frompublicationin printedmediato storytellingforchil- dren"seducation.Likeother formsofartw orks,manyfa- mouscartoonimages werecreatedbased onreal-world scenes.Figure

1showsareal-worldscene whosecor-

?Correspondingauthor (a)Originalsce ne (b)Ourresul t Figure1.An exampleof cartoonstylization.(a) Areal-world scenewhosecorresponding cartoonimageappears intheanimated film“Your Name".(b)Ourresultthattransforms thephoto(a) to thecartoonstyle. Notethatour trainingdatadoes notcontainan y picturein“Y ourName". respondingcartoonimage appearedinthe animatedfilm “YourName".Howev er,manually recreatingreal-world scenesincartoon stylesisv erylaboriousand involv es substantialartisticskills. Toobtain high-qualitycartoons, artistshav etodraweverysingleline andshadeeach color regionoftargetscenes. Meanwhile,existing imageediting software/algorithmswithstandardfeaturescannot produce satisfactoryresultsforcartoonization.Therefore, specially designedtechniquesthat canautomaticallytransform real- worldphotostohigh-qualitycartoon styleimages arevery helpfulandfor artists,tremendousamount oftimecan be savedsothattheycanfocus onmorecreati vew ork.Such toolsalsopro videauseful additiontophotoeditingsoft- waresuchasInstagramand Photoshop.

Stylizingimagesin anartisticmanner hasbeenwidely

studiedinthe domainofnon-photorealistic rendering[ 25].
Traditionalapproachesdevelop dedicatedalgorithmsfor specificstyles.Ho wever ,substantialeffortsarerequired toproducefine-grained stylesthatmimic individualartists. Recently,learning-basedstyletransfermethods (e.g.[ 6]), inwhichan imagecan bestylizedbased onprovidedex- amples,hav edrawnconsiderableattention.Inparticular, thepower ofGenerativeAdversarial Networks(GANs) [ 38]
formulatedina cyclicmanner isexplored toachievehigh- qualitystyletransfer ,withthe distinctfeaturethatthemodel istrainedusing unpairedphotosand stylizedimages.

Althoughsignificantsuccess hasbeenachie vedwith

9465

learningbasedstylization, state-of-the-artmethodsf ailtoproducecartoonizedimages withacceptablequality .Therearetwo reasons.First,insteadofaddingte xturessuchasbrushstrok esinman yotherstyles,cartoonimagesare

highlysimplifiedandabstractedfromreal-world photos. Second,despitev ariationofstyles amongartists,cartoon imageshav enoticeablecommonappearance—clearedges, smoothcolorshading andrelativ elysimplete xtures— whichisv erydifferent fromotherformsofartworks.

Inthispaper ,wepropose CartoonGAN,anovelGAN-

basedapproachto photocartoonization.Our methodtakes asetof photosanda setofcartoon imagesfortraining. To producehighquality resultswhilemaking thetrainingdata easytoobtain, wedonotrequirepairingor correspondence betweentwo setsofimages.Fromthe perspective ofcom- putervisionalgorithms, thegoalof cartoonstylizationis to mapimagesin thephotomanifold intothecartoon mani- foldwhilek eepingthecontent unchanged.Toachieve this goal,wepropose tousea dedicatedGAN-basedarchitec- turetogetherwith twosimple yeteffecti velossfunctions.

Themaincontrib utionsofthis paperare:

fectivelylearnsthemappingfromreal-world photostocar - toonimagesusing unpairedimagesets fortraining.Our methodisable togeneratehigh-quality stylizedcartoons, whicharesubstantially betterthanstate-of-the-art methods. Whencartoonimages fromindividual artistsareused for training,ourmethod isableto reproducetheirstyles. (2)We proposetwosimpleyetef fective lossfunctionsin GAN-basedarchitecture. Inthegenerativenetw ork,tocope withsubstantialstyle variationbetween photosandcar - toons,weintroduce asemantic lossdefinedas an?1sparse regularizationinthehigh-lev elfeaturemaps oftheV GG network[

30].Inthe discriminatornetwork, weproposean

edge-promotingadversarial lossforpreservingclearedges. (3)We furtherintroduceaninitializationphaseto im- provetheconvergenceof thenetwork tothetargetmanifold. Ourmethodis muchmoreef ficienttotrain thanexisting methods.

2.RelatedW ork

2.1.Nonphotorealistic rendering(NPR)

ManyNPRalgorithmshav ebeende veloped,either au-

tomaticallyorsemi-automatically ,tomimic specificartis- ticstylesincluding cartoons[

25].Somew orksrender3D

shapesinsimple shading,whichcreates cartoon-likeef- fect[

28].Suchtechniques calledcelshadingcansav e

substantialamountof timeforartists andhav ebeenused inthecreation ofgames aswellas cartoonvideosand movies[

22].Howe ver,turningexistingphotosorvideos

intocartoonssuch astheproblem studiedinthis paperis muchmorechallenging. Avariet yofmethodshavebeendev elopedtocreate im- ageswithflat shading,mimickingcartoon styles.Such methodsuseeither imagefiltering[

33]orformulations in

optimizationproblems[

35].Howe ver,itisdifficulttocap-

turerichartis ticstylesusing simplemathematicalformulas. Inparticular, applyingfilteringoroptimizationuniformly totheentirely imagedoesnot give thehigh-lev elabstrac- tionthatan artistwould normallydo,such asmakingobject boundariesclear. Toimprovethe results,alternativemeth- odsrelyon segmentationof images/videos[

32],althoughat

odshav ealsobeendevelopedforportraits[

36,26],where

semanticsegmentation canbederivedautomatically byde- tectingfacial components.However, suchmethodscannot copewithgeneral images.

2.2.Stylizationwith neuralnetworks

ConvolutionalNeuralNetworks(CNNs)[

17,18]hav e

receivedconsiderableattentionforsolvingmany computer visionproblems.Instead ofdev elopingspecificNPR al- gorithmswhichrequire substantialeffort foreachstyle, styletransfer hasbeenactivelyresearched. Unliketradi- tionalstyletransfer methods[

11,12]whichrequire paired

style/non-styleimages,recent studies[

19,1,7,8]show that

theVGG network[

30]trainedfor objectrecognitionhas

goodabilityto extractsemantic featuresofobjects, which isvery importantinstylization.Asa result,morepo werful styletransfermethods have beendev elopedwhichdonot requirepairedtraining images.

Givenastyleimageandacontent image,Gatyset al.[

6] firstproposeda neuralstyletransfer (NST)methodbased onCNNsthat transfersthestyle fromthestyle imagetothe contentimage.The yusethe featuremapsofapre-trained

VGGnetworktorepresent thecontentandoptimizethere-

sultimage,such thatitretains thecontentfrom thecontent imagewhilematching thetexture informationofthe style image,wherethe textureis describedusingthe globalGram matrix[

7].Itproduces niceresultsfor transferringav ari-

etyofartistic stylesautomatically. Howev er,it requiresthe contentandstyle imagestobe reasonablysimilar. Further- more,whenimages containmultipleobjects, itmaytransfer stylestosemantically differentre gions.Theresults forcar- toonstyletransfer aremoreproblematic, asthey oftenfail toreproduceclear edgesorsmooth shading.

LiandW and[

20]obtainedstyle transferbylocal match-

forfusion(CNNMRF). Howev er,local matchingcanmake mistakes,resultinginsemantically incorrectoutput.Liao et al.[

21]proposeda DeepAnalogymethod whichkeeps se-

manticallymeaningfuldense correspondencesbetweenthe contentandstyle imageswhiletransferring thestyle.The y alsocompareand blendpatchesin theVGG featurespace.

Chenetal. [

3]proposeda methodtoimpro vecomic style

9466

transferbytraining adedicatedCNN toclassifycomic/non- comicimages.All thesemethodsuse asingle styleimageforacontent image,andthe resultheavily dependsonthe chosenstyleimage, asthereis inevitableambiguity regard- ingtheseparation ofstylesand contentinthe styleimage.Incomparison, ourmethodlearnsacartoonstyle usingtwo setsofimages (i.e.,real-world photosandcartoon images).2.3.Imagesynthesis withGANs

Analternativ e,promisingapproachtoimagesynthesis

istouse Generative AdversarialNetw orks(GANs)[

9,34],

whichproducestate-of-the-art resultsinman yapplications suchastexttoimagetranslation[

24],imageinpainting [37],

imagesuper-resolution [

19],etc.The key ideaofa GAN

modelisto traintwo networks(i.e., ageneratorand adis- criminator)iterativ ely,wherebytheadversariallosspro- videdbythe discriminatorpushesthe generatedimagesto- wardsthetargetmanifold [ 37].

Severalworks[

5,14,16]hav eprovidedGANsolutions

topixel-to-pix elimagesynthesisproblems.However, these methodsrequirepaired imagesetsfor thetrainingprocess whichisimpractical forstylizationdue tothechallenge of obtainingsuchcorresponding imagesets.

Toaddressthisfundamentallimitation, CycleGAN[

38]
wasrecentlyproposed,whichis aframew orkableto per- formimagetranslation withunpairedtraining data.To achievethisgoal,ittrainstwo setsofGAN modelsatthe sametime,mapping fromclassA toclassB andfromclass Btoclass A,respectiv ely.The lossisformulated basedon thecombinedmapping thatmapsimages tothesame class. process.Thismethod alsoproducespoor resultsforcartoon tionandclear edges)ofcartoon images.Asa comparison, ourmethodutilizes aGANmodel tolearnthe mappingbe- tweenphotoand cartoonmanifoldsus ingunpairedtraining data.Thanksto ourdedicatedloss functions,ourmethod is abletosynthesize highqualitycartoon images,andcan be trainedmuchmore efficiently.

2.4.Network architectures

Manyworksshow thatalthoughdeepneuralnetworks

canpotentiallyimpro vethe abilitytorepresentcomplex functions,they canalsobedifficultto trainbecauseof the notoriousvanishing gradientproblem[

29,31].There-

centlyintroducedconcept ofresidualblocks [

10]isa pow-

erfulchoiceto simplifythetraining process.Itdesigns an “identityshortcutconnection" whichreliev esthev anishing gradientissuewhile training.Modelsbased onresidual blockshav eshownimpressiveperformanceingenerati ve networks[

15,19,38].

Anothercommonw aytoease thetrainingofdeepCNNsisbatchnormalization [

13],whichis designedtocounteract

theinternalco variateshift andreducetheoscillationswhen approachingtheminimum point.Inaddition, LeakyReLu (LReLU)[

23]isa widelyusedacti vationfunction indeep

CNNsforef ficientgradientpropag ationwhichincreasesthe performanceofnetw orksbyallo wingasmall,non-zerogra- dientwhenthe unitisnot active. Weinte gratethesetech- niquesinour cartoonizationdeeparchitecture.

3.CartoonGAN

AGANframe workconsists oftwoCNNs.Oneisthe

generatorGwhichis trainedtoproduceoutputthatfools thediscriminator. TheotheristhediscriminatorDwhich classifieswhetherthe imageisfrom therealtar getmani- foldorsynthetic. Wedesign thegeneratorand discrimina- tornetworks tosuittheparticularityofcartoon images;see

Figure

2forano verview .

Weformulatetheprocessof learningtotransform real-

worldphotosintocartoonimages asamapping functionquotesdbs_dbs33.pdfusesText_39