生成对抗网络GANs的用途

超、凢脫俗 2021-12-14 16:27 271阅读 0赞

朋友们，我是床长! 如需转载请标明出处：[http://blog.csdn.net/jiangjunshow][http_blog.csdn.net_jiangjunshow]

简介

如果说目前深度学习最火，应用最多的领域，莫过于 GAN--Generative Adversarial Network，翻译过来就是生成对抗网络，单单从名字上看，你会觉得它就是一个生成模型，看起来就是用于生成图片而已。

实际上，它最开始出现的时候，确实就是用于生成图片，但它可不只是一个生成模型，它实际上是两个网络相互博弈，一个是生成器，也就是生成假图片，另一个就是判别器，用于判断输入图片的真伪，然后目标自然就是让判别器无法判断生成器的图片是真还是假。

当然距离它在 2014 年第一次提出来的时候，已经过去 5 年了，它的应用不仅仅局限在生成图片，越来越多的研究人员把它应用到各个方面，包括图片转换、图像修复、图像超分辨率、风格迁移、文本生成、视频生成等等，今天介绍的这篇文章，就是总结下目前 GANs 可以实现的一些有趣的应用！

文章将这些应用分为以下这些领域，然后会介绍实现该应用的论文，主要是 2016-2018年之间的论文

*  生成图片
 *  人脸生成
 *  照片生成
 *  生成卡通人物
 *  图像转换
 *  文本到图片的转换
 *  语义图片到照片的转换
 *  正脸图片生成
 *  生成新的人体姿势
 *  照片到表情的转换
 *  照片编辑
 *  图片混合
 *  超分辨率
 *  图片修复
 *  衣服转换
 *  视频预测
 *  3D 物体生成

--------------------

1. 生成图片

这是 2014 年最早提出 GANs 的论文 “[Generative Adversarial Networks][]” 中所实现的应用，如下图所示，包括生成 MNIST 手写数字数据集、CIFAR10 小物体图片、人脸数据集的图片。

![1][]

接着 2015 年的论文[Unsupervised Representation Learning with Deep Convolutional Generative Adversarial Networks][]，也被称为 **DCGAN** 实现了稳定使用 CNN 训练 GAN ，其结果如下图所示：

![1 1][]

2. 人脸生成

人脸方面的应用本来就是计算机视觉领域最热门也是应用最深、技术最成熟的其中一个方向，GANs 自然也涉及到这方面的应用了。

2017 年的论文 "[Progressive Growing of GANs for Improved Quality, Stability, and Variation][Progressive Growing of GANs for Improved Quality_ Stability_ and Variation] "，简称 `ProGAN`， 可以做到生成非常逼真的人脸，如下图所示

![1 2][]

这篇论文还展示了它的其他应用，生成其他物体的实验结果：

![1 3][]

另外，2018 年的一份报告 “[The Malicious Use of Artificial Intelligence: Forecasting, Prevention, and Mitigation][The Malicious Use of Artificial Intelligence_ Forecasting_ Prevention_ and Mitigation]” 描述了从 2014 年到 2017 年 GANs 的快速发展，并且以人脸生成作为例子，如下展示这几年人脸生成的结果的变化，确实是越来越逼真了。

![1 4][]

3. 照片生成

2018 年的论文“[Large Scale GAN Training for High Fidelity Natural Image Synthesis][]” ，也叫作 `BigGAN` ,在生成真实照片方面做出非常好的结果，如下图所示，当初发表的时候，也是引起很大的关注--[学界 | 史上最强GAN图像生成器，Inception分数提高两倍][_ _GAN_Inception].

![1 5][]

4. 生成卡通人物

2017年的论文 “[Towards the Automatic Anime Characters Creation with Generative Adversarial Networks][]” 则是将 GANs 应用到生成日本动漫人物的人脸方面的应用了，如下图所示

![1 6][]

此外也有人应用 GANs 生成宠物小精灵的图片，如下图所示，其项目地址为：

*  [github.com/moxiegushi/…][github.com_moxiegushi]
 *  [github.com/kvpratama/g…][github.com_kvpratama_g]

![1 7][]

不过最近也有人用 GANs 来生成不同属性的神奇宝贝：

[利用CycleGAN生成不同属性的神奇宝贝][CycleGAN]

![1 8][]

5. 图像转换

图像转换是将 GANs 应用在很多转换的任务上，这里最著名的一篇论文就是2016年的 “[Image-to-Image Translation with Conditional Adversarial Networks][]” ，也就是 `pix2pixGAN`，它可以实现这些图片的转换：

*  将语义图片转换为街景和建筑的照片
 *  卫星照片转成谷歌地图
 *  照片从白天转为夜晚的景色
 *  黑白照片上色
 *  素描图转彩色图片

下面是论文的展示结果，第一行分别就是语义图片转街景、语义图片转建筑图片、黑白图片上色，第二行就是卫星照片转谷歌地图、白天转为夜晚以及素描图片转彩色图片。

![1 9][]

但 `pix2pixGAN` 对数据集要求是成对，即输入图片和其期望输出图片是一对，但这对数据集要求很高，很多时候并没有这样成对的图片，于是 2017 年有了一篇改进的论文 “[Unpaired Image-to-Image Translation using Cycle-Consistent Adversarial Networks][]”，也就是 `CycleGAN` ，它只需要原始领域和目标领域的数据集即可，不需要一一对应的成对数据，它可以实现以下几种转换：

*  照片转为艺术画风格
 *  普通的马和斑马的转换
 *  照片从夏天变为冬天的风格
 *  卫星图片转谷歌地图

其实现结果如下所示，第一行就是艺术画和照片转换、斑马和普通马的转换、夏天和冬天季节转换，而第二行、第三行则是具体介绍了每种转换的一个例子。

![1 10][]

6. 文本到图片的转换

2016 年的一篇论文 “[StackGAN: Text to Photo-realistic Image Synthesis with Stacked Generative Adversarial Networks][StackGAN_ Text to Photo-realistic Image Synthesis with Stacked Generative Adversarial Networks]” ，介绍了采用 `StackGAN` 来实现通过简单的对如鸟类和花朵的文本描述，生成逼真的照片。如下图展示了两个例子，两句话的生成结果，第一句话是描述的是一个头部为红色，然后羽毛从头到尾是逐渐从红色渐变为灰色的鸟，而第二句话描述的是深绿色并有一个短嘴的鸟。

![1 11][]

2016年的另外一篇论文 “[Generative Adversarial Text to Image Synthesis][]” 则可以实现更多的文本到图片的描述，包括生成鸟类、花朵等等，如下图所示：

![1 12][]

其他相似的论文还有：

*  [TAC-GAN – Text Conditioned Auxiliary Classifier Generative Adversarial Network][TAC-GAN _ Text Conditioned Auxiliary Classifier Generative Adversarial Network]，2017
 *  [Learning What and Where to Draw][]，2016

7. 语义图片到照片的转换

2017年的论文 “[High-Resolution Image Synthesis and Semantic Manipulation with Conditional GANs][]” 采用了条件 GANs 方法来生成非常逼真的照片，它可以根据给定的语义照片生成对应不同类型的照片：

*  街景照片
 *  卧室照片
 *  人脸照片
 *  给定素描图片生成人脸照片

一个生成街景照片的例子如下图所示：

![1 13][]

8. 正脸图片生成

2017 年的论文 “[Beyond Face Rotation: Global and Local Perception GAN for Photorealistic and Identity Preserving Frontal View Synthesis][Beyond Face Rotation_ Global and Local Perception GAN for Photorealistic and Identity Preserving Frontal View Synthesis]” 实现了给定非正脸的输入照片，生成正脸的照片结果。这个可以应用在对人脸验证或者人脸识别系统中。

效果如下图所示：

![1 14][]

9. 生成新的人体姿势

2017 年论文 “[Pose Guided Person Image Generation][]” 实现了可以给定输入图片，然后生成的姿势，如下图所示，输入是正向，侧面或者背面姿势，都可以生成新的姿势，包括正向的生成侧面图片等等；

![1 15][]

10. 照片到表情的转换

2016 年的论文--“[Unsupervised Cross-Domain Image Generation][]” 使用 GAN 来生成不同领域的图片，比如街景数量到手写字体数据集，然后再生成哪种程度的表情或者是卡通人物的脸。如下所示：

![1 16][]

11. 照片编辑

CVPR 2018 的论文 “[StarGAN: Unified Generative Adversarial Networks for Multi-Domain Image-to-Image Translation][StarGAN_ Unified Generative Adversarial Networks for Multi-Domain Image-to-Image Translation]" 实现了对照片编辑，主要是对人脸属性的编辑，如下图所示，它可以修改人脸的一些属性，包括头发颜色、表情、性别、年龄变化等，这都取决于训练集是否包含对应的标签。

![1 17][]

starGAN 已经开源，项目地址是：

[github.com/yunjey/star…][github.com_yunjey_star]

其他相似的论文有：

*  [Invertible Conditional GANs For Image Editing][]，2016
 *  [Coupled Generative Adversarial Networks][]，2016
 *  [Neural Photo Editing with Introspective Adversarial Networks][]，2016
 *  [Image De-raining Using a Conditional Generative Adversarial Network][]，2017

下面几篇主要是针对人脸年龄变化：

*  [Face Aging With Conditional Generative Adversarial Networks][]，2017
 *  [Age Progression/Regression by Conditional Adversarial Autoencoder][Age Progression_Regression by Conditional Adversarial Autoencoder]，2017

12. 图片混合

2017年的论文 [GP-GAN: Towards Realistic High-Resolution Image Blending][GP-GAN_ Towards Realistic High-Resolution Image Blending] 采用 GANs 来实现图片的混合操作，即融合多张图片的不同元素，如下图所示，它是将图 a 中间部分融合到图 b 同样位置。

![1 18][]

13. 超分辨率

图像超分辨率技术指的是根据**低分辨率图像生成高分辨率图像**的过程，该技术希望根据已有的图像信息重构出**缺失的图像细节**。

ECCV 2018 的论文--[ESRGAN: Enhanced super-resolution generative adversarial networks][ESRGAN_ Enhanced super-resolution generative adversarial networks] 提出的 ESRGAN，即增强型超分辨率生成对抗网络，它可以将真实的细节添加到低分辨率的图像中，因此产生更精细的画面。其实现的结果如下所示：

![\*马克思·佩恩原版游戏截图与使用 ESRGAN 超分辨率重制游戏的截图。\*][ESRGAN]

ESRGAN 的项目地址：

[github.com/xinntao/ESR…][github.com_xinntao_ESR]

不仅可以实现对图片的超分辨率，对视频的超分辨率也有人采用 GANs 技术进行实现--[Temporally Coherent GANs for Video Super-Resolution (TecoGAN)][Temporally Coherent GANs for Video Super-Resolution _TecoGAN]，这篇论文首次提出了一种对抗和循环训练方法，以监督空间高频细节和时间关系。具体介绍可以查看下面这篇文章的介绍：

[低清视频也能快速转高清：超分辨率算法TecoGAN][TecoGAN]

其他实现超分辨率的论文有：

*  [Photo-Realistic Single Image Super-Resolution Using a Generative Adversarial Network][]，2016
 *  [High-Quality Face Image SR Using Conditional Generative Adversarial Networks][]，2017
 *  [Analyzing Perception-Distortion Tradeoff using Enhanced Perceptual Super-resolution Network][]，2018

14. 图片修复

2019 年的论文 [EdgeConnect: Generative Image Inpainting with Adversarial Edge Learning][EdgeConnect_ Generative Image Inpainting with Adversarial Edge Learning] 对图片的修复分为两步，边缘生成然后进行图像补全，具体介绍可以看下：

[女神被打码了？一笔一划脑补回来，效果超越Adobe | 已开源][Adobe _]

其效果如下，分别展示了六个例子，图 a 是需要修复的图片，图 b 就是中间生成的边缘图，图 c 是最终修复的结果。

![1 19][]

项目地址：

[github.com/knazeri/edg…][github.com_knazeri_edg]

其他论文有：

*  [Image Inpainting via Generative Multi-column Convolutional Neural Networks][]，2018
 *  [Generative Image Inpainting with Contextual Attention][]， 2018
 *  [High-resolution image inpainting using multi-scale neural patch synthesis][]，CVPR 2017
 *  [Generative Face Completion][]，2017
 *  [Context Encoders: Feature Learning by Inpainting][Context Encoders_ Feature Learning by Inpainting]，2016

15. 2d试衣

2017 年的论文--[The Conditional Analogy GAN: Swapping Fashion Articles on People Images][The Conditional Analogy GAN_ Swapping Fashion Articles on People Images]，尝试采用 GANs 实现 2d 试衣的效果，论文给出结果如下，它是给定一个模特和对应需要更换的衣服，然后实现替换模特身上的衣服。

![1 20][]

国外有人根据这篇文章进行一些修改，写了篇博客介绍，并且开源了其代码，其结果如下所示：

博客：[shaoanlu.wordpress.com/2017/10/26/…][shaoanlu.wordpress.com_2017_10_26]

Github 地址：[github.com/shaoanlu/Co…][github.com_shaoanlu_Co]

![1 21][]

目前来看，这个技术还不是非常成熟。

其他相似的论文：

*  [INSTAGAN][]，2018，Github：[github.com/sangwoomo/i…][github.com_sangwoomo_i]

16. 视频预测

2016 年的论文--[Generating Videos with Scene Dynamics][] 介绍了如何用 GANs 实现视频预测，主要是应用于静态场景里面的元素，如下图所示：

![1 22][]

17. 3D 物体生成

2016 年的论文--[Learning a Probabilistic Latent Space of Object Shapes via 3D Generative-Adversarial Modeling][] 介绍了如何通过 GAN 生成新的三维物体，比如椅子、车、沙发、桌子等等，如下图所示：

![1 23][]

2016年的另一篇论文--[3D Shape Induction from 2D Views of Multiple Objects][] 也同样实现给定一张多个视角的二维物体图片，生成三维物体，如下图所示：

![1 24][]

--------------------

小结

更多的关于 GANs 的应用，还可以阅读下面的文章和 Github 项目

*  [gans-awesome-applications: Curated list of awesome GAN applications and demo][gans-awesome-applications_ Curated list of awesome GAN applications and demo].
 *  [Some cool applications of GANs][], 2018.
 *  [GANs beyond generation: 7 alternative use cases][GANs beyond generation_ 7 alternative use cases], 2018.

[http_blog.csdn.net_jiangjunshow]: http://blog.csdn.net/jiangjunshow
[Generative Adversarial Networks]: https://link.juejin.im?target=https%3A%2F%2Farxiv.org%2Fabs%2F1406.2661
[1]: https://user-gold-cdn.xitu.io/2019/6/29/16ba3879bf35fb0a?imageView2/0/w/1280/h/960/format/webp/ignore-error/1
[Unsupervised Representation Learning with Deep Convolutional Generative Adversarial Networks]: https://link.juejin.im?target=https%3A%2F%2Farxiv.org%2Fabs%2F1511.06434
[1 1]: https://user-gold-cdn.xitu.io/2019/6/29/16ba3879bee5b279?imageView2/0/w/1280/h/960/format/webp/ignore-error/1
[Progressive Growing of GANs for Improved Quality_ Stability_ and Variation]: https://link.juejin.im?target=https%3A%2F%2Farxiv.org%2Fabs%2F1710.10196
[1 2]: https://user-gold-cdn.xitu.io/2019/6/29/16ba3879bf81a3a6?imageView2/0/w/1280/h/960/format/webp/ignore-error/1
[1 3]: https://user-gold-cdn.xitu.io/2019/6/29/16ba3879c10dc9ed?imageView2/0/w/1280/h/960/format/webp/ignore-error/1
[The Malicious Use of Artificial Intelligence_ Forecasting_ Prevention_ and Mitigation]: https://link.juejin.im?target=https%3A%2F%2Farxiv.org%2Fabs%2F1802.07228
[1 4]: https://user-gold-cdn.xitu.io/2019/6/29/16ba3879c0b9fb14?imageView2/0/w/1280/h/960/format/webp/ignore-error/1
[Large Scale GAN Training for High Fidelity Natural Image Synthesis]: https://link.juejin.im?target=https%3A%2F%2Farxiv.org%2Fabs%2F1809.11096
[_ _GAN_Inception]: https://link.juejin.im?target=https%3A%2F%2Fmp.weixin.qq.com%2Fs%3F__biz%3DMzA3MzI4MjgzMw%3D%3D%26mid%3D2650749368%26idx%3D2%26sn%3D4b970da824cc7c6fb0fa3014315da7fa%26scene%3D0%23wechat_redirect
[1 5]: https://user-gold-cdn.xitu.io/2019/6/29/16ba3879c174209a?imageView2/0/w/1280/h/960/format/webp/ignore-error/1
[Towards the Automatic Anime Characters Creation with Generative Adversarial Networks]: https://link.juejin.im?target=https%3A%2F%2Farxiv.org%2Fabs%2F1708.05509
[1 6]: https://user-gold-cdn.xitu.io/2019/6/29/16ba3879fa4e5f9c?imageView2/0/w/1280/h/960/format/webp/ignore-error/1
[github.com_moxiegushi]: https://link.juejin.im?target=https%3A%2F%2Fgithub.com%2Fmoxiegushi%2FpokeGAN
[github.com_kvpratama_g]: https://link.juejin.im?target=https%3A%2F%2Fgithub.com%2Fkvpratama%2Fgan%2Ftree%2Fmaster%2Fpokemon
[1 7]: https://user-gold-cdn.xitu.io/2019/6/29/16ba387a035829d9?imageView2/0/w/1280/h/960/format/webp/ignore-error/1
[CycleGAN]: https://link.juejin.im?target=https%3A%2F%2Fmp.weixin.qq.com%2Fs%2F0xn57qy2CQiUvbF_wWwrlw
[1 8]: https://user-gold-cdn.xitu.io/2019/6/29/16ba387a02846521?imageView2/0/w/1280/h/960/format/webp/ignore-error/1
[Image-to-Image Translation with Conditional Adversarial Networks]: https://link.juejin.im?target=https%3A%2F%2Farxiv.org%2Fabs%2F1611.07004
[1 9]: https://user-gold-cdn.xitu.io/2019/6/29/16ba3879fd3810bf?imageView2/0/w/1280/h/960/format/webp/ignore-error/1
[Unpaired Image-to-Image Translation using Cycle-Consistent Adversarial Networks]: https://link.juejin.im?target=https%3A%2F%2Farxiv.org%2Fabs%2F1703.10593
[1 10]: https://user-gold-cdn.xitu.io/2019/6/29/16ba387a031be608?imageView2/0/w/1280/h/960/format/webp/ignore-error/1
[StackGAN_ Text to Photo-realistic Image Synthesis with Stacked Generative Adversarial Networks]: https://link.juejin.im?target=https%3A%2F%2Farxiv.org%2Fabs%2F1612.03242
[1 11]: https://user-gold-cdn.xitu.io/2019/6/29/16ba387a0cd2c758?imageView2/0/w/1280/h/960/format/webp/ignore-error/1
[Generative Adversarial Text to Image Synthesis]: https://link.juejin.im?target=https%3A%2F%2Farxiv.org%2Fabs%2F1605.05396
[1 12]: https://user-gold-cdn.xitu.io/2019/6/29/16ba387a26b3952e?imageView2/0/w/1280/h/960/format/webp/ignore-error/1
[TAC-GAN _ Text Conditioned Auxiliary Classifier Generative Adversarial Network]: https://link.juejin.im?target=https%3A%2F%2Farxiv.org%2Fabs%2F1703.06412
[Learning What and Where to Draw]: https://link.juejin.im?target=https%3A%2F%2Farxiv.org%2Fabs%2F1610.02454
[High-Resolution Image Synthesis and Semantic Manipulation with Conditional GANs]: https://link.juejin.im?target=https%3A%2F%2Farxiv.org%2Fabs%2F1711.11585
[1 13]: https://user-gold-cdn.xitu.io/2019/6/29/16ba387a2a40b67e?imageView2/0/w/1280/h/960/format/webp/ignore-error/1
[Beyond Face Rotation_ Global and Local Perception GAN for Photorealistic and Identity Preserving Frontal View Synthesis]: https://link.juejin.im?target=https%3A%2F%2Farxiv.org%2Fabs%2F1704.04086
[1 14]: https://user-gold-cdn.xitu.io/2019/6/29/16ba387a358cbfd8?imageView2/0/w/1280/h/960/format/webp/ignore-error/1
[Pose Guided Person Image Generation]: https://link.juejin.im?target=https%3A%2F%2Farxiv.org%2Fabs%2F1705.09368
[1 15]: https://user-gold-cdn.xitu.io/2019/6/29/16ba387a52c290d7?imageView2/0/w/1280/h/960/format/webp/ignore-error/1
[Unsupervised Cross-Domain Image Generation]: https://link.juejin.im?target=https%3A%2F%2Farxiv.org%2Fabs%2F1611.02200
[1 16]: https://user-gold-cdn.xitu.io/2019/6/29/16ba387a52b46d54?imageView2/0/w/1280/h/960/format/webp/ignore-error/1
[StarGAN_ Unified Generative Adversarial Networks for Multi-Domain Image-to-Image Translation]: https://link.juejin.im?target=https%3A%2F%2Farxiv.org%2Fabs%2F1711.09020
[1 17]: https://user-gold-cdn.xitu.io/2019/6/29/16ba387a52c8fd21?imageView2/0/w/1280/h/960/format/webp/ignore-error/1
[github.com_yunjey_star]: https://link.juejin.im?target=https%3A%2F%2Fgithub.com%2Fyunjey%2Fstargan
[Invertible Conditional GANs For Image Editing]: https://link.juejin.im?target=https%3A%2F%2Farxiv.org%2Fabs%2F1611.06355
[Coupled Generative Adversarial Networks]: https://link.juejin.im?target=https%3A%2F%2Farxiv.org%2Fabs%2F1606.07536
[Neural Photo Editing with Introspective Adversarial Networks]: https://link.juejin.im?target=https%3A%2F%2Farxiv.org%2Fabs%2F1609.07093
[Image De-raining Using a Conditional Generative Adversarial Network]: https://link.juejin.im?target=https%3A%2F%2Farxiv.org%2Fabs%2F1701.05957
[Face Aging With Conditional Generative Adversarial Networks]: https://link.juejin.im?target=https%3A%2F%2Fieeexplore.ieee.org%2Fdocument%2F8296650
[Age Progression_Regression by Conditional Adversarial Autoencoder]: https://link.juejin.im?target=https%3A%2F%2Farxiv.org%2Fabs%2F1702.08423
[GP-GAN_ Towards Realistic High-Resolution Image Blending]: https://link.juejin.im?target=https%3A%2F%2Farxiv.org%2Fabs%2F1703.07195
[1 18]: https://user-gold-cdn.xitu.io/2019/6/29/16ba387baea5a92a?imageView2/0/w/1280/h/960/format/webp/ignore-error/1
[ESRGAN_ Enhanced super-resolution generative adversarial networks]: https://link.juejin.im?target=https%3A%2F%2Farxiv.org%2Fabs%2F1809.00219
[ESRGAN]: https://user-gold-cdn.xitu.io/2019/6/29/16ba387a7cde6c59?imageView2/0/w/1280/h/960/format/webp/ignore-error/1
[github.com_xinntao_ESR]: https://link.juejin.im?target=https%3A%2F%2Fgithub.com%2Fxinntao%2FESRGAN
[Temporally Coherent GANs for Video Super-Resolution _TecoGAN]: https://link.juejin.im?target=https%3A%2F%2Farxiv.org%2Fpdf%2F1811.09393.pdf
[TecoGAN]: https://link.juejin.im?target=https%3A%2F%2Fwww.jiqizhixin.com%2Farticles%2F2019-04-16-9%3Ffrom%3Dsynced%26keyword%3D%25E8%25B6%2585%25E5%2588%2586%25E8%25BE%25A8%25E7%258E%2587
[Photo-Realistic Single Image Super-Resolution Using a Generative Adversarial Network]: https://link.juejin.im?target=https%3A%2F%2Farxiv.org%2Fabs%2F1609.04802
[High-Quality Face Image SR Using Conditional Generative Adversarial Networks]: https://link.juejin.im?target=https%3A%2F%2Farxiv.org%2Fabs%2F1707.00737
[Analyzing Perception-Distortion Tradeoff using Enhanced Perceptual Super-resolution Network]: https://link.juejin.im?target=https%3A%2F%2Farxiv.org%2Fabs%2F1811.00344
[EdgeConnect_ Generative Image Inpainting with Adversarial Edge Learning]: https://link.juejin.im?target=https%3A%2F%2Farxiv.org%2Fabs%2F1901.00212
[Adobe _]: https://link.juejin.im?target=https%3A%2F%2Fmp.weixin.qq.com%2Fs%2FF8o_zBBvuWyW90uyP5bLvQ
[1 19]: https://user-gold-cdn.xitu.io/2019/6/29/16ba387a95d4f1c1?imageView2/0/w/1280/h/960/format/webp/ignore-error/1
[github.com_knazeri_edg]: https://link.juejin.im?target=https%3A%2F%2Fgithub.com%2Fknazeri%2Fedge-connect
[Image Inpainting via Generative Multi-column Convolutional Neural Networks]: https://link.juejin.im?target=https%3A%2F%2Farxiv.org%2Fabs%2F1810.08771
[Generative Image Inpainting with Contextual Attention]: https://link.juejin.im?target=https%3A%2F%2Farxiv.org%2Fabs%2F1801.07892
[High-resolution image inpainting using multi-scale neural patch synthesis]: https://link.juejin.im?target=https%3A%2F%2Farxiv.org%2Fabs%2F1611.09969
[Generative Face Completion]: https://link.juejin.im?target=https%3A%2F%2Farxiv.org%2Fabs%2F1704.05838
[Context Encoders_ Feature Learning by Inpainting]: https://link.juejin.im?target=https%3A%2F%2Farxiv.org%2Fabs%2F1604.07379
[The Conditional Analogy GAN_ Swapping Fashion Articles on People Images]: https://link.juejin.im?target=https%3A%2F%2Farxiv.org%2Fabs%2F1709.04695
[1 20]: https://user-gold-cdn.xitu.io/2019/6/29/16ba387a737243a1?imageView2/0/w/1280/h/960/format/webp/ignore-error/1
[shaoanlu.wordpress.com_2017_10_26]: https://link.juejin.im?target=https%3A%2F%2Fshaoanlu.wordpress.com%2F2017%2F10%2F26%2Freimplement-conditional-anology-gan-in-keras%2F
[github.com_shaoanlu_Co]: https://link.juejin.im?target=https%3A%2F%2Fgithub.com%2Fshaoanlu%2FConditional-Analogy-GAN-keras
[1 21]: https://user-gold-cdn.xitu.io/2019/6/29/16ba387a9c2cb655?imageView2/0/w/1280/h/960/format/webp/ignore-error/1
[INSTAGAN]: https://link.juejin.im?target=https%3A%2F%2Farxiv.org%2Fpdf%2F1812.10889.pdf
[github.com_sangwoomo_i]: https://link.juejin.im?target=https%3A%2F%2Fgithub.com%2Fsangwoomo%2Finstagan
[Generating Videos with Scene Dynamics]: https://link.juejin.im?target=https%3A%2F%2Farxiv.org%2Fabs%2F1609.02612
[1 22]: https://user-gold-cdn.xitu.io/2019/6/29/16ba387ab1182206?imageView2/0/w/1280/h/960/format/webp/ignore-error/1
[Learning a Probabilistic Latent Space of Object Shapes via 3D Generative-Adversarial Modeling]: https://link.juejin.im?target=https%3A%2F%2Farxiv.org%2Fabs%2F1610.07584
[1 23]: https://user-gold-cdn.xitu.io/2019/6/29/16ba387ac3124b5b?imageView2/0/w/1280/h/960/format/webp/ignore-error/1
[3D Shape Induction from 2D Views of Multiple Objects]: https://link.juejin.im?target=https%3A%2F%2Farxiv.org%2Fabs%2F1612.05872
[1 24]: https://user-gold-cdn.xitu.io/2019/6/29/16ba387ac7fbedde?imageView2/0/w/1280/h/960/format/webp/ignore-error/1
[gans-awesome-applications_ Curated list of awesome GAN applications and demo]: https://link.juejin.im?target=https%3A%2F%2Fgithub.com%2Fnashory%2Fgans-awesome-applications
[Some cool applications of GANs]: https://link.juejin.im?target=https%3A%2F%2Fmedium.com%2F%40jonathan_hui%2Fgan-some-cool-applications-of-gans-4c9ecca35900
[GANs beyond generation_ 7 alternative use cases]: https://link.juejin.im?target=https%3A%2F%2Fmedium.com%2F%40alexrachnog%2Fgans-beyond-generation-7-alternative-use-cases-725c60ba95e8

生成对抗网络GANs的用途

发表评论取消回复

还没有评论，来说两句吧...

相关阅读