引用本文:王星,杜伟,陈吉,等.基于深度残差生成式对抗网络的样本生成方法[J].控制与决策,2020,35(8):1887-1894
【打印本页】   【HTML】   【下载PDF全文】   查看/发表评论  【EndNote】   【RefMan】   【BibTex】
←前一篇|后一篇→ 过刊浏览    高级检索
本文已被:浏览次   下载 本文二维码信息
码上扫一扫!
分享到: 微信 更多
基于深度残差生成式对抗网络的样本生成方法
王星,杜伟,陈吉,陈海涛
(辽宁工程技术大学电子与信息工程学院,辽宁葫芦岛125105)
摘要:
作为样本生成的重要方法之一,生成式对抗网络(GAN)可以根据任意给定数据集中的数据分布生成样本,但它在实际的训练过程中存在生成样本纹理模糊、训练过程不稳定以及模式坍塌等问题.针对以上问题,在深度卷积生成式对抗网络(DCGAN)的基础上,结合残差网络,设计一种基于深度残差生成式对抗网络的样本生成方法RGAN.该样本生成方法利用残差网络和卷积网络分别构建生成模型和判别模型,并结合正负样本融合训练的学习优化策略进行优化训练.其中:深度残差网络可以恢复出丰富的图像纹理;正负样本融合训练的方式可以增加对抗网络的鲁棒性,有效缓解对抗网络训练不稳定和模式坍塌现象的发生.在102 Category Flower Dataset数据集上设计多个仿真实验,实验结果表明RGAN能有效提高生成样本的质量.
关键词:  生成式对抗网络  残差网络  深度学习  对抗训练  RGAN  FID
DOI:10.13195/j.kzyjc.2018.1700
分类号:TP183
基金项目:国家自然科学基金项目(61402212);辽宁省高等学校杰出青年学者成长计划项目(LJQ2015045);中国博士后基金面上项目(2016M591452);辽宁省自然科学基金面上项目(2015020098).
Sample generation based on residual generative adversarial network
WANG Xing,DU Wei,CHEN Ji,CHEN Hai-tao
(School of Electronic and Information Engineering,Liaoning Technical University,Huludao 125105,China)
Abstract:
As one of the important methods of sample generation, the generative adversarial network(GAN) can generate samples based on the data distribution in any given data set, but it has generated sample texture blur, unstable training process and mode collapse in the actual training process. In order to solve the above problems, this paper proposes a residual generative adversarial network,(RGAN) based on the deep convolutional generative adversarial network(DCGAN) and the residual network. The sample generation method uses the residual network and the convolution network to construct the generator and the discriminator respectively, and combines the learning optimization strategy of the positive and negative sample fusion training to optimize the training process. Among them, the depth residual network can recover the rich image texture and, the positive and negative sample fusion training can increase the robustness against the network and effectively alleviate the instability of network training and the collapse of the model. This paper designs several simulation experiments on the 102 category flower dataset. The experimental results show that the RGAN can effectively improve the quality of generated samples.
Key words:  generative adversarial network  residual network  deep learning  adversarial training  residual generative adversarial network  FID

用微信扫一扫

用微信扫一扫