基于半监督学习的多层感知器生成对抗网络  

Multilayer Perceptron Generative Adversarial Network Based on Semi-supervised Learning

作　　者：王格格 郭涛[1] 余游 苏菡[1] WANG Ge-ge;GUO Tao;YU You;SU Han(Department of Computer Science,Sichuan Normal University,Chengdu 610101,China)

机构地区：[1]四川师范大学计算机科学学院

出　　处：《小型微型计算机系统》2019年第11期2297-2303,共7页Journal of Chinese Computer Systems

基　　金：国家自然科学基金项目(61403266)资助;四川省可视化与虚拟现实重点实验室项目(KJ201419)资助

摘　　要：半监督学习通过充分利用大量无标记数据和少量有标记数据来改善学习性能,近年来已成为机器学习领域的研究热点.半监督生成对抗网络SGAN将生成对抗网络扩展到半监督学习,通过在原始无标记输入数据的基础上加入少量有标记数据,并将判别器转换成分类器输出分类结果,以此来解决传统分类问题中因有标记训练数据太少引起的过拟合问题.但SGAN判别器上的线性卷积层提取图像深层次特征的能力较弱,使其在半监督环境下对图像进行分类的准确率不高,且生成的图像质量较差.为此,提出半监督多层感知器生成对抗网络SMPGAN.该网络采用多层感知器卷积层代替SGAN判别器上的线性卷积层来提高抽象层次,并在生成器上使用特征匹配进一步提高图像的分类精度.在不同数量的有标记样本辅助下,SMPGAN的分类精度和图像生成效果均有明显提升.Semi-supervised learning,which improves learning performance by making full use of a large amount of unlabeled data and a small amount of labeled data,becomes a research hotspot in the field of machine learning in recent years. The semi-supervised generative adversarial network( SGAN) extends the generative adversarial network to semi-supervised learning. By adding a small amount of labeled data on the basis of original unlabeled input data and converting the discriminator into a classifier to output classification results,the overfitting problem caused by few labeled training data in traditional classification problems is solved. However,the ability of the linear convolutional layer on the SGAN discriminator to extract the deep features of the image is weak. It makes the classification accuracy of the image in the semi-supervised environment not high,and also makes the quality of the generated image poor. For this reason,a semi-supervised multilayer perceptron generative adversarial network( SMPGAN) is proposed. The network uses a multilayer perceptron convolutional layer instead of the linear convolutional layer on the SGAN discriminator to improve the level of abstraction,and uses feature-matching on the generator to further improve the classification accuracy of the image. With the help of different numbers of labeled samples,the classification accuracy and image generation effect of SMPGAN are significantly improved.

关 键 词：半监督学习 生成对抗网络 多层感知器 特征匹配 

分 类 号：TP181[自动化与计算机技术—控制理论与控制工程]