ROBUST DEEP LEARNING FOR NON-BLIND IMAGE DECONVOLUTION

Abstract

Non-blind image deconvolution is an ill-posed linear inverse problem with a wide range of applications. Recently, deep learning has been a prominent tool for solving such a problem with promising performance. However, existing deep-learning-based approaches are sensitive to both noise level uncertainty and kernel/model error, which limits their wider adoption in practice. This dissertation presents two robust deep learning solutions to address these sensitivities. For handling the noise level uncertainty, we proposed a deep neural network that unrolls the variational expectation-maximization (VEM) algorithms. For boosting the robustness to kernel/model error, we proposed another network inspired by the error-in-variables (EIV) model with the total-least-squares (TLS) solver. Extensive experiments showed that the proposed methods achieved state-of-the-art results.