Abstract

Semi-supervised learning, which entails training a model with manually labeled images and pseudo-labels for unlabeled images, has garnered considerable attention for its potential to improve image classification performance. Nevertheless, incorrect decision boundaries of classifiers and wrong pseudo-labels for beneficial unlabeled images below the confidence threshold increase the generalization error in semi-supervised learning. This study proposes a novel framework for semi-supervised learning termed consistency-regularized bad generative adversarial network (CRBSGAN) through a new loss function. The proposed model comprises a discriminator, a bad generator, and a classifier that employs data augmentation and consistency regularization. Local augmentation is created to compensate for data scarcity and boost bad generators. Moreover, label consistency regularization is considered for bad fake images, real labeled images, unlabeled images, and latent space for the discriminator and bad generator. In the adversarial game between the discriminator and the bad generator, feature space is better captured under these conditions. Furthermore, local consistency regularization for good-augmented images applied to the classifier strengthens the bad generator in the generator–classifier adversarial game. The consistency-regularized bad generator produces informative fake images similar to the support vectors located near the correct classification boundary. In addition, the pseudo-label error is reduced for low-confidence unlabeled images used in training. The proposed method reduces the state-of-the-art error rate from 6.44 to 4.02 on CIFAR-10, 2.06 to 1.56 on MNIST, and 6.07 to 3.26 on SVHN using 4000, 3000, and 500 labeled training images, respectively. Furthermore, it achieves a reduction in the error rate on the CINIC-10 dataset from 19.38 to 15.32 and on the STL-10 dataset from 27 to 16.34 when utilizing 1000 and 500 labeled images per class, respectively. Experimental results and visual synthesis indicate that the CRBSGAN algorithm is more efficient than the methods proposed in previous works. The source code is available at https://github.com/ms-iraji/CRBSGAN ↗.

中文翻译：

---

具有一致性正则化不良半监督生成对抗网络的图像分类：视觉数据分析和合成

摘要

半监督学习需要使用手动标记的图像和未标记图像的伪标签来训练模型，因其提高图像分类性能的潜力而受到广泛关注。然而，分类器的错误决策边界和低于置信阈值的有益未标记图像的错误伪标签会增加半监督学习中的泛化误差。这项研究通过新的损失函数提出了一种新的半监督学习框架，称为一致性正则化不良生成对抗网络（CRBSGAN）。所提出的模型包括鉴别器、不良生成器和采用数据增强和一致性正则化的分类器。创建局部增强是为了补偿数据稀缺并增强不良生成器。此外，对于不良假图像、真实标记图像、未标记图像以及鉴别器和不良生成器的潜在空间，考虑了标签一致性正则化。在鉴别器和不良生成器之间的对抗性游戏中，在这些条件下可以更好地捕获特征空间。此外，应用于分类器的良好增强图像的局部一致性正则化增强了生成器-分类器对抗游戏中的不良生成器。一致性正则化不良生成器生成与位于正确分类边界附近的支持向量类似的信息丰富的假图像。此外，对于训练中使用的低置信度未标记图像，伪标签误差也减少了。该方法分别使用 4000、3000 和 500 个标记训练图像，将 CIFAR-10 上的最新错误率从 6.44 降低到 4.02，MNIST 上从 2.06 降低到 1.56，SVHN 上从 6.07 降低到 3.26。此外，当每类使用 1000 个和 500 个标记图像时，它分别将 CINIC-10 数据集上的错误率从 19.38 降低到 15.32，将 STL-10 数据集上的错误率从 27 降低到 16.34。实验结果和视觉综合表明，CRBSGAN 算法比以前的工作中提出的方法更有效。源代码可在 https://github.com/ms-iraji/CRBSGAN ↗ 获取。