生成式对抗网络研究与应用进展

Research and Application Progress of Generative Adversarial Networks

基于零和博弈思想的生成式对抗网络(GAN)可通过无监督学习获得数据的分布,并生成较逼真的数据。基于GAN的基础概念及理论框架,研究各类GAN模型及其在特定领域的应用情况,从数据相似性度量、模型框架、训练方法3个方面进行分析,对GAN改进与扩展的相关研究成果进行总结,并从图像合成、风格迁移等应用领域展开讨论,归纳出GAN的优势与不足,同时对其应用前景进行展望。分析结果表明,GAN的学习能力与可塑性强,改进潜力大,应用范围广,但其发展面临的挑战是训练过程不稳定,且缺乏生成数据质量的客观评价标准。

The Generative Adversarial Networks(GAN) based on the zero-sum game idea can obtain the distribution of data through unsupervised learning and generate more realistic data.Based on the basic concepts and theoretical framework of the generated confrontation network,the GAN models and the application results in specific fields are studied,and the data similarity measure,model framework and training method are summarized.The research and related research results of the improvement and expansion are analyzed,and the practical application fields such as image synthesis and style migration are discussed.The advantages and disadvantages of GAN are summarized,and the application prospects are prospected.Analysis results show that the GAN has strong learning ability and plasticity,great potential for improvement and wide application range.However,its development challenges are unstable training process and lack of objective evaluation criteria for generating data quality.