Deep Equilibrium Architectures for Inverse Problems in Imaging

Recent efforts on solving inverse problems in imaging via deep neural networks use architectures inspired by a fixed number of iterations an optimization method. The is typically quite small due to difficulties training corresponding more iterations; the resulting solvers cannot be run for at test time without incurring significant errors. This paper describes alternative approach infinite iterations, yielding consistent improvement reconstruction accuracy above state-of-the-art alternatives and where computational budget can selected optimize context-dependent trade-offs between computation. proposed leverages ideas from Deep Equilibrium Models, fixed-point iteration constructed incorporate known forward model insights classical optimization-based methods.