生成对抗网络图像处理综述

A survey on generative adversarial networks in image processing

2014年提出的生成对抗网络(Generative Adversarial Networks,GAN)是近年来神经网络领域中为数不多的一项新锐技术。GAN在常见生成模型的基础上增加了一个判别模型,以形成巧妙的对抗学习机制,使它能够产生更高质量的图像。近年来各种改进型GAN在图像处理领域得到广泛应用,不但覆盖了几乎所有传统图像处理领域,还包括一些新应用,如图像编辑、图像翻译、风格转移等,普遍取得了胜过传统方法的良好结果。文中在简要分析GAN的系统结构、对抗生成和网络训练的基础上,重点介绍了为提高GAN性能、克服现存缺陷和满足不同应用而出现的多种改进型GAN,如DC-GAN、W-GAN、Big-GAN等。尽管如此,目前GAN尚处于初始发展阶段,将来的前途不可估量。

Generative adversarial networks(GAN)proposed in 2014 is one cutting-edge technique of several deep learning technologies in the neural networks in the past few years.A discriminative model is added to familiar generative model by GAN to form an adversarial competitive strategy in subtle way in order to produce higher quality images.Recently,many improved GANs are widely applied in image processing,not only covering nearly all applications of traditional image processing,but also including other new applications,such as image editing,image translation,style transfer,etc,and have gotten better performance than that the traditional methods got.Based on brief analysis on GAN system architecture,adversarial generating and networks training,this paper stresses on introducing several schemes of improved GANs,such as DC-GAN,W-GAN,Big-GAN,and so on,whose target is to promote GAN’s performance,overcome GAN’s defects,and adapt the environment requirements for deferent applications.Nowadays,GAN techniques are still in initial stage of vigorous development,and have a bright future for us.