作者投稿和查稿 主编审稿 专家审稿 编委审稿 远程编辑

计算机工程 ›› 2019, Vol. 45 ›› Issue (5): 155-160,168. doi: 10.19678/j.issn.1000-3428.0050529

• 人工智能及识别技术 • 上一篇    下一篇

基于分段损失的生成对抗网络

刘其开1,姜代红2,李文吉3   

  1. 1.中国矿业大学 信息与控制工程学院,江苏 徐州 221116; 2.徐州工程学院 信电工程学院,江苏 徐州 221111; 3.中国国土资源航空物探遥感中心 国土资源部航空地球物理与遥感地质重点实验室,北京 100083
  • 收稿日期:2018-02-27 出版日期:2019-05-15 发布日期:2019-05-15
  • 作者简介:刘其开(1992—),男,硕士,主研方向为深度学习、计算机视觉;姜代红,教授、博士;李文吉,工程师、硕士。
  • 基金资助:

    国家自然科学基金(51574232);国土资源部航空地球物理与遥感地质重点实验室航遥青年创新基金(2016YFL02);徐州市科技计划项目(KC16SQ78)。

Generative Adversarial Network Based on Piecewise Loss

LIU Qikai1,JIANG Daihong2,LI Wenji3   

  1. 1.School of Information and Control Engineering,China University of Mining and Technology,Xuzhou,Jiangsu 221116,China; 2.School of Information and Electrical Engineering,Xuzhou University of Technology,Xuzhou,Jiangsu 221111,China; 3.Key Laboratory of Airborne Geophysics and Remote Sensing Geology,Ministry of Land and Resources,China Aerospace Geophysical Survey and Remote Sensing Center for Land and Resources,Beijing 100083,China
  • Received:2018-02-27 Online:2019-05-15 Published:2019-05-15

摘要:

生成对抗网络(GAN)在训练过程中未能有效进行生成器与鉴别器间的同步更新,导致模型训练不稳定并出现模式崩溃的现象。为此,提出一种基于分段损失的生成对抗网络PL-GAN。生成器在不同的训练时期采用不同形式的损失函数,同时引入真实样本与生成样本之间的特征级损失,从而使鉴别器提取的特征更具有鲁棒性。MNIST和CIFAR-10数据集上的实验结果表明,与regular GAN、feature-wise GAN相比,PL-GAN具有更高的分类精度与运行效率。

关键词: 生成对抗网络, 模式崩溃, 特征级损失, 分段损失, 半监督学习

Abstract:

Generative Adversarial Network(GAN) fails to effectively execute the synchronous update between generator and discriminator during training,resulting in unstable model training and mode collapse.To solve this problem,a generative adversarial network PL-GAN based on piecewise loss is proposed.The generator uses different loss functions in different training periods,and introduces the feature-wise loss between the real sample and the generated sample,which makes the feature extracted by the discriminator more robust.Experimental results on MNIST and CIFAR-10 datasets show that PL-GAN has higher classification accuracy and operation efficiency than regular GAN and feature-wise GAN.

Key words: Generative Adversarial Network(GAN), mode collapse, feature-wise loss, piecewise loss, Semi-Supervised Learning(SSL)

中图分类号: