生成式对抗网络模型研究

Survey of Generative Adversarial Networks Models

在系统的总结GAN原始模型的提出背景、基本原理与基本框架的基础上,归纳总结了基于距离度量与能量模型角度而提出的衍进模型f-GAN、WGAN、WGAN-GP、EBGAN;针对解决原始GAN模型的不稳定性而提出的衍进模型DCGAN、Improved GAN、PGGAN;基于模型结合角度而提出的GAN+LAP、GAN+LSTM、GAN+CVAE、GAN+AE以及针对增强模型实用性而提出的衍进模型SGAN、CGAN、InfoGAN。对GAN的一些具体应用领域和场景进行了梳理和介绍。

Based on the background, basic principles and basic framework of the GAN original model, the evolution models f-GAN, WGAN, WGAN-GP and EBGAN based on the distance metric and energy model were summarized;the evolution model DCGAN, Improved GAN and PG-GAN which were proposed to solve the instability of the original GAN model were summarized;GAN + LAP, GAN-FLSTM, GAN + CVAE and GAN+AE were summarized based on the model combination angle;and the evolution models SGAN, CGAN, and InfoGAN which were proposed for the enhancement of the model's practicality were summarized. Some specific application areas and scenarios of GAN had been reviewed and introduced.