Proposal of New Activation Function in Deep Image Prior

摘要

We propose a new activation function and verify its performance in a deep image prior network. The new activation function is RSwish, which is Swish with a random slope forx < 0 assuming that the input value of the activation function isx. In addition, we manipulate the probability of the random number and observe the effect. As a result, we found that RSwish can perform better than Swish by manipulating probabilities according to the degree of color change in the image. (c) 2020 Institute of Electrical Engineers of Japan. Published by Wiley Periodicals LLC.