Abstract: Aims: We consider the problem of image blind deconvolution through a variational Bayesian algorithm. Methods: We introduce our algorithm VBA (variational Bayesian algorithm) accounting efficiently for a wide range of priors on the unknowns. Then, VBA iterations are embedded into a deep neural network architecture, following the recently introduced deep unrolling paradigm. Results: Our experiments illustrate the excellent performance of our new method on two datasets, comprising grayscale and color images, and degraded with various kernel types. Compared to state-of-the-art variational and deep learning approaches, our method delivers a more accurate estimation of both the image and the blur kernels. It also includes an automatic noise estimation step, so that it requires little hyperparameter tuning. The proposed method is very competitive in terms of computational time during the test phase, while showing similar train time to its deep learning competitors. Conclusions: This talk presents our novel method [1] for blind image deconvolution that combines a variational Bayesian algorithm with a neural network architecture. The main core of the proposed architecture is highly interpretable, as it implements unrolled iterates of a well sounded Bayesian-based blind deconvolution method. Acknowledgements: The authors acknowledge support from the Agence Nationale de la Recherche of France under MAJIC (ANR-17-CE40-0004-01) project, and from the European Research Council Starting Grant MAJORIS ERC-2019- STG-850925.
[1] Y. Huang, E. Chouzenoux, and J.-C. Pesquet. Unrolled Variational Bayesian Algorithm for Image Blind Deconvolution. Tech. Rep., 2021. https://arxiv.org/abs/2110.07202

Biography: Emilie Chouzenoux (IEEE Senior Member, 2020) received the engineering degree from Ecole Centrale, Nantes, France, in 2007, and the Ph.D. degree in signal processing from the Institut de Recherche en Communications et Cybern\’etique (IRCCyN, UMR CNRS 6597), Nantes, in 2010. Between 2011 and 2019, she was a Ma\^itre de conferences at the University of Paris-Est Marne-la-Vall\’ee, Champs-sur-Marne, France (LIGM, UMR CNRS 8049). Since September 2019, she has been a Researcher at Inria Saclay, in CVN lab at CentraleSup\’elec, University Paris Saclay, France. She is an Associated Editor of IEEE Transactions in Signal Processing, and of SIAM Journal on Mathematics of Data Science. Since January 2020, she has been the PI of the ERC Starting Grant MAJORIS. Her research interests are in large scale optimization algorithms for inverse problems and machine learning problems of image processing.