Generative adversarial network (GAN) has been successfully extended to solve semi-supervised image classification tasks recently. However, it is still a great challenge for GAN to exploit the unlabeled images for boosting its classification ability when labeled images are very limited. In this paper, we propose a novel CCS-GAN model for semi-supervised image classification, which aims to improve its classification ability by utilizing the cluster structure of unlabeled images and ’bad’ generated images. Specifically, it employs a new cluster consistency loss to constrain its classifier to keep the local discriminative consistency in each cluster of unlabeled images and thus provides implicit supervised information to boost the classifier. Meanwhile, it adopts an enhanced feature matching approach to encourage its generator to produce adversarial images from the low-density regions of real distribution, which can enhance the discriminative ability of the classifier during adversarial training and suppress the mode collapse problem. Extensive experiments on four benchmark datasets show that: the proposed CCS-GAN achieves very competitive performance in semi-supervised image classification tasks when compared with several state-of-the-art competitors.

最近，生成对抗网络 (GAN) 已成功扩展到解决半监督图像分类任务。然而，当标记图像非常有限时，GAN 仍然是一个巨大的挑战，它利用未标记的图像来提高其分类能力。在本文中，我们提出了一种用于半监督图像分类的新型 CCS-GAN 模型，旨在通过利用未标记图像和“坏”生成图像的聚类结构来提高其分类能力。具体来说，它采用新的集群一致性损失来约束其分类器，以保持每个未标记图像集群的局部判别一致性，从而提供隐式监督信息来提升分类器。同时，它采用增强的特征匹配方法来鼓励其生成器从真实分布的低密度区域生成对抗性图像，这可以增强分类器在对抗性训练期间的判别能力并抑制模式崩溃问题。在四个基准数据集上的大量实验表明：与几个最先进的竞争对手相比，所提出的 CCS-GAN 在半监督图像分类任务中取得了非常有竞争力的性能。