Deep image prior for inverse problems: acceleration and probabilistic treatment

Abstract

Since its first proposal in 2018, deep image prior has emerged as a very powerful unsupervised deep learning technique for solving inverse problems. The approach has demonstrated very encouraging empirical success in image denoising, deblurring, super-resolution etc. However, there are also several known drawbacks of the approach, notably high computational expense. In this talk, we describe some of our efforts: we propose to accelerate the training process by pretraining on synthetic dataset and further we propose a novel probabilistic treatment of deep image prior to facilitate uncertainty quantification.

Speaker

Bangti Jin's research interests include inverse problems, numerical analysis and machine learning. Currently he serves on the editorial board of five journals, including Inverse Problems and Journal of Computational Mathematics.

https://www.math.cuhk.edu.hk/~btjin/