Speaker
Description
Image restoration methods often suffer from the "Regression to the Mean" (R2M) effect, leading to blurry results due to their inability to restore high-frequency details. This is problematic in microscopy, where the loss of such fine details can deter subsequent analysis and downstream processing.
In this work, we propose to tackle this challenge through a data-driven approach, by leveraging the iterative prediction power in denoising diffusion models. Diffusion models, with their forward process and iterative restoration, have a higher likelihood of recovering a plausible mode for the degraded image. This works especially well, when the forward operator is a known function. However, in some cases, we do not have a controllable forward operator. In this work, we will discuss the application of diffusion models in microscopy image restoration tasks for both known and unknown forward operators to iteratively recover the desired image.
Additionally, being a generative model, our technique allows for sampling multiple plausible restored solutions, in contrast to traditional methods that predict only the MMSE estimate. In this work, we explore the challenges of quantifying and leveraging the posterior distribution of our approach and present preliminary results of our method.
| Authors | Anirban Ray*, Ashesh Ashesh, Florian Jug |
|---|---|
| Keywords | Diffusion Models, Widefield Microscopy, Image Restoration, Generative Models |
