Title:	Image Denoising with Stochastic Differential Equations
DNr:	Berzelius-2024-443
Project Type:	LiU Berzelius
Principal Investigator:	Ziwei Luo <ziwei.luo@it.uu.se>
Affiliation:	Uppsala universitet
Duration:	2024-12-01 – 2025-06-01
Classification:	10207
Homepage:	https://algolzw.github.io/
Keywords:

Abstract

The Wiener process is to add continuous Gaussian noises to a clean signal, which constructs the dispersion part of a Stochastic Differential Equation (SDE). In this project, we model the image noise-adding process as a conditional Wiener process where the noisy image is an intermediate state and the transition follows a simple SDE. Naturally, we can reverse the SDE to get the initial state image. Its drift function of the reverse-time SDE can be derived from the original forward SDE with an additional score function which can be approximated with a deep convolutional neural network (CNN). We would like to evaluate this approach on various of tasks like unconditional/conditional image generation, image restoration, and image translation. This project aims to breakthrough the current frameworks of image generation, leading to a better result in realistic image/signal restoration. This project has produced 5 top AI conference papers (at CVPR, ICML, ICLR, and NeurIPS) and all of these papers have acknowledged Berzelius, KAW, and NSC.