Improved SinGAN's Performance by Changing the Activation Function. Issue 2 (18th October 2021)

• Record Type: Journal Article
• Authors: Segawa, Ryo; Hayashi, Hitoshi
• Is Part Of: IEEJ transactions on electrical and electronic engineering. Volume 17:Issue 2(2022)
• Page Start: 308
• Page End: 310
• Publication Date: 2021-10-18
• Subjects: deep learning -- GANs -- unsupervised learning -- activation function -- super‐resolution; Electrical engineering -- Periodicals; Electronics -- Periodicals; 621.3
• DOI: 10.1002/tee.23514
• Languages: English
• ISSNs: 1931-4973

Abstract

Abstract : Generative adversarial nets (GANs) perform well on a variety of tasks, but rely on large datasets and expensive computer‐based learning. SinGAN was proposed as a GANs that overcomes this problem, but its super‐resolution performance for large‐scale natural images was not very good. In this study, we aimed to improve the performance of SinGAN by changing the activation function. As a result of the verification, it was found that applying RSwish to the generator and Swish to the discriminator can generate an image with less deterioration as a whole than the default. © 2021 Institute of Electrical Engineers of Japan. Published by Wiley Periodicals LLC.