Disclosed herein is a method for performing blind deconvolutions of blurred images. The method approximates the proximal operators for the data fidelity term and the prior term of a minimization function using trained neural networks and solves the minimization using iterations of the Douglas-Rachford algorithm.