基于半监督编码生成对抗网络的图像分类模型

在实际应用中,为分类模型提供大量的人工标签越来越困难,因此,近几年基于半监督的图像分类问题获得了越来越多的关注.而大量实验表明,在生成对抗网络(Generative adversarial network,GANs)的训练过程中,引入少量的标签数据能获得更好的分类效果,但在该类模型的框架中并没有考虑用于提取图像特征的结构,为了进一步利用其模型的学习能力,本文提出一种新的半监督分类模型.该模型在原生成对抗网络模型中添加了一个编码器结构,用于直接提取图像特征,并构造了一种新的半监督训练方式,获得了突出的分类效果.本模型分别在标准的手写体识别数据库MNIST、街牌号数据库SVHN和自然图像数据库CIFAR-10上完成了数值实验,并与其他半监督模型进行了对比,结果表明本文所提模型在使用少量带标数据情况下得到了更高的分类精度.

A Semi-supervised Encoder Generative Adversarial Networks Model for Image Classification

The semi-supervised image classification task has attracted more and more attention recently owing to the problem that adequate labeled data is hard to acquire from industrial applications. Meanwhile, considerable works demonstrate that the improved generative adversarial networks (GANs) can achieve great classification performance with only few labeled images. Intuitively, GAN is a generative model, there is no semantic feature extractor in the main framework. In order to further utilize the ability of GANs, we propose to add an encoder in the framework to extract features of images directly, and simultaneously to use a new semi-supervised training method to train this new image classification model. The classification results of experiments have shown the state-of-the-art accuracy performance in semi-supervised MNIST, SVHN and CIFAR-10.