预览加载中,请您耐心等待几秒...
1/10
2/10
3/10
4/10
5/10
6/10
7/10
8/10
9/10
10/10

亲,该文档总共36页,到这已经超出免费预览范围,如果喜欢就直接下载吧~

如果您无法下载资料,请参考说明:

1、部分资料下载需要金币,请确保您的账户上有足够的金币

2、已购买过的文档,再次下载不重复扣费

3、资料包下载后请先用软件解压,在使用对应软件打开

基于全卷积神经网络的图像修复算法研究161203105333·2020的基于全卷积神经网络的图像修复算法研究的研究的的的2020基于全卷积神经网络的图像修复算法研究图像修复算的研究的图像修复算法基于图像积的的图像修复的的基于的图像修复算法于的修复的网络的基于全卷积神经网络的图像修复算法神经网络的网络图像经全全图像图像全网络的网络全的修复的图像图像的的的图像的图像CelebA算法修复图像修复全卷积神经网络神经网络2020ResearchonimageinpaintingbasedonFullyConvolutionalNetworksAbstractImageinpaintingisadifficultprobleminthefieldofdeeplearning.Traditionalimageinpaintingalgorithmsaremostlybasedonimagediffusionorpatch,andusuallyrequiresmoothnessprior.Thisalgorithmsoftenperformpoorlyinsceneswithlargemissingareasorstrongsemanticscenes.Thedevelopmentofdeeplearningtechnologyhasmadeimagerepairhaveanewresearchdirection.Imageinpaintingalgorithmsbasedondeeplearningcanbeappliedtoawiderrangeofscenes,greatlyimprovingtherepaireffect.However,thedepthgenerationmodelstillhasgreatroomforimprovement,suchassinglenetworkstructure,unstabletrainingprocess,poorinterpretability,andlackoftheoreticalcertificationforsuperparameterselection.ThispaperbringforwardanimageinpaintingalgorithmbasedonFCN.WithreferencetotheideaofGAN,thenetworkstructureofthecontextencoderisimproved,sothatthepartiallymissingimageisvisuallyconsistentwiththeglobalimageaftercompletion.Inordertotraintheconsistencyoftheimagecompletionnetwork,thediscriminantnetworkisdividedintotwoparts:theglobaldiscriminatortakesthecompleteimageastheinputtoidentifytheglobalconsistencyofthescene;Thelocaldiscriminatoronlyobservesth