Self-supervised Bayesian Deep Learning for Image Denoising

Tongyao Pang, Yuhui Quan, Hui Ji

Keywords: Self-supervised Learning, Bayesian Inference, Image Denoising

Abstract:

Deep learning is currently one prominent approach for image denoising, and most of existing works train a denoising neural network (NN) on many pairs of noisy images and their clean counterparts. Recent studies showed that it is possible to train a denoising NN on a dataset consisting of only noisy images. This paper took one step further to study how to train a powerful denoising NN for a given image without any training samples, which is appealing to the applications where collecting training samples is challenging. For instance, biological imaging and medical imaging.
Built on the Bayesian neural network (BNN), this paper proposed a self-supervised deep learning method for denoising a single image, in the absence of training samples. The experiments showed that the performance of our self-supervised method is very competitive to those state-of-the-art supervised ones.

Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics

Reviewed Version (pdf): https://openreview.net/references/pdf?id=iuqdHjH3b