CSE Faculty Candidate Seminar - Yu Sun

Name: Yu Sun, postdoctoral scholar at California Institute of Technology

Date: Tuesday, March 5, 2024 at 11:00 am

Location: Coda Building, Second Floor, Room 230 (Google Maps link)

Link: The recording of this in-person seminar will be uploaded to CSE's MediaSpace

Coffee, drinks, and snacks provided!

Title: Turning Denoisers into Principled Imaging Solvers: Algorithm, Theory, and Application

Abstract: Many imaging systems suffer from information loss during the data acquisition. This makes the problem of image reconstruction an ill-posed inverse problem, where the ill-posedness can ultimately lead to strong image artifacts and multiple solutions that equally explain the measurements. This talk will present how a simple image denoiser, whose goal is to simply remove Gaussian noise from an image, can become the building block for designing advanced algorithms for solving these challenges. The talk will focus on two works. In the first work, we will introduce how the image denoiser can help us design a principled algorithm that can accurately explore the full solution space of ill-posed imaging inverse problems while also impose expressive image priors. Here, the image denoiser serves as a bridge to incorporate the recent advances in the score-based generative models for generating different images. Additionally, we propose a theoretical analysis that can characterize the convergence behavior of our algorithm under nonlinear inverse problems and imperfect denoiser. We demonstrate our method on the challenging black hole imaging problem where multi-modal solutions exist. The second work is focused on a specific 3D microscopy application, where traditional image reconstruction approaches lead to significant artifacts. In this work, the image denoiser becomes a key component in learning a deep neural-field imaging solver for removing heavy artifacts. Here, the image denoiser functions as an indispensable image regularizer.

Bio: Dr. Yu Sun currently works as a postdoc scholar at California Institute of Technology (Caltech) with Prof. Katie Bouman. His research is focused on developing principled computational imaging algorithms integrated with machine learning models. Prior to Caltech, He received the Ph.D. degree from Washington University in St. Louis (Wash U) supervised by Prof. Ulugbek S. Kamilov. His dissertation was the winner of the 2022 Turner Ph.D. Dissertation Award (top in the class) at Wash U. He is an elected member of the IEEE Signal Processing Society Technical Committee of Computational Imaging (CI TC) and a consultant associated editor of IEEE Open Journal of Signal Processing. His work has been published in prestigious journals and conferences including Nature Machine Intelligence, NeurIPS, and ICLR.