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

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