Comparison of Performance by Activation Functions on Deep Image Prior

Abstract

In this paper, we compare the performance of activation functions on a deep image prior. The activation functions considered here are the standard rectified linear unit (ReLU), leaky rectified linear unit (Leaky ReLU), and the randomized leaky rectified linear unit (RReLU). We use these functions for denoising, super-resolution, and inpainting of the deep image prior. Our aim is to observe the effect of differences in the activation functions.