Generative Models for Blind Inverse Problems in Imaging
Brett Levac (PhD student at UT Austin)
BRETT LEVAC
Title: Generative Models for Blind Inverse Problems in Imaging
Abstract: Often in computational imaging there is uncertainty in the forward process which was used to collect measurements. End-to-End techniques can accommodate this by pairwise training, however, their performance can suffer in the presence of distribution shifts at test time. Model-based recovery techniques, on the other hand, require us to solve for both the unknowns in the imaging system and the desired image. We demonstrate how generative models can be used for model-based signal recovery in blind inverse problems, specifically in motion corrupted MRI. Additionally, we propose a technique leveraging generative models that can be used to identify distributions of corrupted forward models from unpaired training data. We demonstrate that these learned priors can later be leveraged to solve blind inverse problems.
Website: https://scholar.google.com/citations?hl=en&user=rRHEl5UAAAAJ&view_op=list_works&sortby=pubdate
In Room M7 101, 1st floor, Monod campus, ENSL.