计算机工程与应用 ›› 2021, Vol. 57 ›› Issue (8): 26-35.DOI: 10.3778/j.issn.1002-8331.2011-0322

• 热点与综述 • 上一篇    下一篇

生成对抗网络及其图像处理应用研究进展

王晋宇,杨海涛,李高源,张长弓,冯博迪   

  1. 1.航天工程大学 研究生院,北京 101416
    2.航天工程大学 航天信息学院,北京 101416
  • 出版日期:2021-04-15 发布日期:2021-04-23

Research Progress of Generative Adversarial Network and Its Application in Image Processing

WANG Jinyu, YANG Haitao, LI Gaoyuan, ZHANG Changgong, FENG Bodi   

  1. 1.Graduate School, Space Engineering University, Beijing 101416, China
    2.School of Space Information, Space Engineering University, Beijing 101416, China
  • Online:2021-04-15 Published:2021-04-23

摘要:

生成对抗网络(GAN)是一种基于对抗思想的架构体系。作为人工智能大发展背景下诞生的前沿算法,GAN已经在图像处理的多个领域取得了显著的成果。从传统GAN的算法入手,对其模型架构、数学机理、优缺点进行剖析。总结了具有代表性的GAN变体,并对GAN在图像处理方面的前沿应用进行介绍。结合现有GAN发展依然存在的问题,对GAN的发展趋势进行了展望。

关键词: 生成对抗网络, 纳什均衡, 结构变体, 损失变体

Abstract:

Generative Adversarial Network(GAN) is an architecture based on the idea of confrontation. As a cutting-edge algorithm born under the background of the great development of artificial intelligence, GAN has made remarkable achievements in many fields of image processing. Starting with the traditional GAN algorithm, the model structure, mathematical mechanism, advantages and disadvantages of GAN are analyzed. The representative GAN variants are summarized, and the frontier applications of GAN in image processing are introduced. Combined with the existing problems in the development of GAN, the development trend of GAN is prospected.

Key words: Generative Adversarial Network(GAN), Nash equilibrium, structure variant, loss variant