Extension of Image Data Using Generative Adversarial Networks and Application to Identification of Aurora

Abstract

In recent years, automatic auroral image classification has been actively investigated. The baseline method has relied on supervised learning. As this approach requires a large amount of labeled teacher data, it is necessary to collect the data manually and label them, which is a time-consuming task. In this study, we proposed a method to extend an image data set by inputting training images into a deep convolutional generative adversarial network (DCGAN) and generating images in this manner. The proposed approach implied using both generated and original images to train the classifier. It could reduce the number of labeling operations performed manually. As an evaluation experiment, we performed classifier learning on the data sets before and after extension and confirmed that the classification accuracy was improved because of training on the data set after the extension.