Data Augmentation For Deep Learning Using Generative Adversarial Networks
Daiki Yorioka, Hyunho Kang, Keiichi Iwamura
Abstract
Deep learning requires the use of several labeled images as training data. However, in practice, it is difficult to obtain a sufficient number of appropriate images, and it is particularly difficult to obtain diverse classes of images with labels. This poses a challenge in the training of convolutional neural networks (CNNs). Data augmentation, a technique for generating more data from pre-existing training data, plays an essential role in addressing the dearth of appropriate images. Thus, in this study, we propose a data augmentation method using image generation and a generative adversarial network (GAN) with geometric deformation. In the proposed method, a small dataset is repeatedly augmented via geometric deformation and used as training data for the auxiliary classifier-GAN. Finally, the performance of the CNN trained on the dataset generated by the proposed method is evaluated. Although the quality and accuracy of the results achieved were insufficient, it would be remiss to dismiss the potential of this approach. The challenge now lies in improving the quality of the generated images, thereby improving the accuracy of the CNN.