Presentation information

General Session

General Session » GS-2 Machine learning

[4G2-GS-2k] 機械学習:基礎理論

Fri. Jun 11, 2021 11:00 AM - 12:40 PM Room G (GS room 2)

座長:谷口 忠大(立命館大学)

12:20 PM - 12:40 PM

[4G2-GS-2k-05] Homogeneous responsive activation function "Yamatani Activation" and application to single-image super-resolution

〇Takuma Yoshimura1 (1. poco-apoco Networks Co. Ltd.)

Keywords:Single image super resolution, Homogeneous responsiveness

In this study, we propose Yamatani Activation, which is an activation function satisfying homogeneity with two variables as input, and show its application to single image super-resolution. A neural network consisting only of Yamatani Convolution without a bias term using Yamatani Activation responds to the input in such a way that the output satisfies the homogeneity.
When this is used for single-image super-resolution, it is possible to achieve super-resolution that is independent of the dynamic range of the image and that responds homogeneously to differential components. In order to verify the usefulness of this property, we created artificial images with various edge patterns, trained a neural network with these images, and verified whether this neural network can also super-resolve real images. As a result, PSNR:32.56, SSIM:0.915 were achieved in RGB 3-channel evaluation condition, DIV2K, x2.

Authentication for paper PDF access

A password is required to view paper PDFs. If you are a registered participant, please log on the site from Participant Log In.
You could view the PDF with entering the PDF viewing password bellow.