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