生成式对抗网络的应用综述
A Review of Application of Generative Adversarial Networks
投稿时间:2019-05-20  修订日期:2019-09-05
DOI:10.11908/j.issn.0253-374x.19204     稿件编号:    中图分类号:TP181
 
摘要点击次数: 83    全文下载次数: 50
中文摘要
      生成式对抗网络(GAN)是一种优秀的生成式模型,能够不依赖任何先验假设,学习到高维复杂的数据分布。这一强大的性能使得它成为近年来研究的热点,并在诸多应用领域取得了显著的研究成果。首先介绍了生成式对抗网络的基本原理,各种目标函数以及常用的模型结构。然后,详细分析了生成式对抗网络在条件限制下生成图片的各种演进方法。此外,介绍了生成式对抗网络在不同领域的应用,包括高分辨率图像生成、小目标检测、非图像数据生成、医学图像分割等方面的最新研究进展。最后,总结了生成式对抗网络训练过程中的优化技巧。旨在通俗地阐明GAN的基础理论以及发展历程,并从应用角度对未来工作进行了展望。
英文摘要
      Generative adversarial networks (GAN) is an excellent generative model, which can learn high-dimensional and complex real data distribution without relying on any prior assumptions. This powerful performance makes it a research hotspot in recent years, and remarkable progress has been made in research in many application fields. In this paper, the basic principle of the GAN, various objective functions and common model structures are introduced. Then, the evolutional methods for generating images under the constraints of conditional generative adversarial networks are analyzed in detail. After that, the applications of the GAN in different fields are introduced, including high-resolution image generation, small target detection, non-image data generation, medical image segmentation and so on. Finally, the optimization techniques in the training process of the GAN are summarized. The purpose of this paper is to elucidate the basic theory and development history of GAN, and to forcast the future work from the perspective of application.
HTML   查看全文  查看/发表评论  

您是第6894340位访问者
版权所有《同济大学学报(自然科学版)》
主管单位:教育部 主办单位:同济大学
地  址: 上海市四平路1239号 邮编:200092 电话:021-65982344 E-mail: zrxb@tongji.edu.cn
本系统由北京勤云科技发展有限公司设计