2020 Volume 8 Issue 2 Pages 110-120
In training deep neural networks for supervised learning tasks, we often use data augmentation methods to increase training dataset sizes. Furthermore, this technique is particularly useful when the size of the training datasets is small, such as when the content of the training datasets includes privacy issues that cannot be made public, or the categories in the train datasets are unbalanced. During the training process, small training datasets will lead to model overfitting. Nowadays, data augmentation methods using Generative Adversarial Network (GAN) and Neural Style showed to provide performance improvements for the task of supervised learning. However, the traditional GAN is easy to cause the network to collapse,which makes the generation process free and uncontrollable. Hence, it may cause the network model to fail to produce deterministic results, which makes the application limited. In this paper, we propose an improved GAN-based data augmentation method for image classification tasks. We compare our model with the latest GAN model, and the results show that our algorithm is effective. On generated images,when applying synthetic images to facial expression attribute classification task, our method achieves 72.5% accuracy rate on the FER2013 PrivateTest datasets and 71.2% accuracy rate on the FER2013 PublicTest datasets.