Self-Supervised Image Denoising with Swin Transformer

Pavel Chizhov, Mikhail Papkov

Keywords: self-supervised image denoising, vision transformers

TL;DR: Testing Swin Transformer-based backbone inside of two frameworks for self-supervised image denoising instead of U-Net.

Abstract: Self-supervised image denoising aims to reconstruct signal from a noisy image with no additional information. Typically, this is accomplished by means of specific frameworks built upon fully-convolutional neural networks. In two such frameworks, Noise2Self and Noise2Same, we replaced conventional convolutional backbones with a state-of-the-art Swin Transformer-based model. In this paper, we summarize the results of experiments on a range of datasets and examine the advantages and limitations of transformers in self-supervised denoising.