首页 | 本学科首页   官方微博 | 高级检索  
     

生成对抗网络的发展与挑战
引用本文:董永生,范世朝,张宇,马尽文. 生成对抗网络的发展与挑战[J]. 信号处理, 2023, 39(1): 154-175. DOI: 10.16798/j.issn.1003-0530.2023.01.015
作者姓名:董永生  范世朝  张宇  马尽文
作者单位:1.河南科技大学信息工程学院,河南 洛阳 471023
基金项目:河南省重点研发与推广专项192102210121
摘    要:生成对抗网络(Generative adversarial network, GAN)由生成模型和判别模型构成,生成模型获取真实数据的概率分布,判别模型判断输入是真实数据还是生成器生成的数据,二者通过相互对抗训练,最终使生成模型学习到真实数据的分布,使判别模型无法准确判断输入数据的来源。生成对抗网络为视觉分类任务的算法性能的提升开辟了新的思路,自诞生之日起至今已经在各个领域产生了大量变体。本文的主要内容包括:生成对抗网络的研究现状、应用场景和基本模型架构,并列举了生成对抗网络本身所存在的弊端;从网络架构、损失函数和训练方式这三方面对生成对抗网络的各种主要典型发展进行归纳;详细总结和分析了生成对抗网络在人脸图像生成和编辑、风格迁移、图像超分辨率、图像修复,序列数据生成、视频生成等各个应用领域的算法以及对应算法的优缺点;介绍了生成对抗网络的常用评价指标并且分析了这些指标的适用场景和不足之处;最后从多个方面对生成对抗网络所面临的挑战进行了讨论,并指出了对其可能的改进方向。

关 键 词:生成对抗网络  生成模型  概率分布估计  应用场景  模型评价
收稿时间:2022-11-30

Development and Challenge of Generative Adversarial Network
Affiliation:1.College of Information Engineering, Henan University of Science and Technology, Luoyang, Henan 471023, China2.School of Mathematical Sciences, Peking University, Beijing 100871, China
Abstract:? ?Generative adversarial network is composed of generative model and discriminant model. Generative model mainly obtains the probability distribution of real data, and the discriminant model is used to judge whether the input is real data or the data generated by the generator. Through the training of the two models against each other, the generative model can finally learn the distribution of real data completely. It also makes it impossible for the discriminant model to accurately determine whether the input data comes from the generating model or the discriminant model. At the same time, it opened up a new way of thinking for the improvement of algorithm performance in visual classification tasks. Since its inception, there have been a lot of variations in a variety of related fields. The main contents of this paper include: (1) We gave a brief introduction about the current research status of generative adversarial network and application scenarios, and further describe the basic model architecture and the drawbacks of generative adversarial network itself. (2) We presented the development and improved methods of generative adversarial network from three aspects: network architecture, loss function and training mode. (3) We typically described a variety of applications of generative adversarial network, as well as the advantages and disadvantages of these algorithms. These typical applications included face image generation and editing, style transfer, image super resolution, image restoration, sequence data generation, video generation and other application fields. (4) The common evaluation indexes for generative adversarial networks were introduced and the applicable scenarios and shortcomings of these indexes were analyzed. (5) Finally, we discussed and presented the challenges of generative adversarial network from several representative aspects, and further we pointed out the possible improvement directions in the future. 
Keywords:
点击此处可从《信号处理》浏览原始摘要信息
点击此处可从《信号处理》下载免费的PDF全文
设为首页 | 免责声明 | 关于勤云 | 加入收藏

Copyright©北京勤云科技发展有限公司  京ICP备09084417号