Energy Dissipation with Plug-and-Play Priors

Abstract

Neural networks have reached outstanding performance for solving various ill-posed inverse problems in imaging. However, drawbacks of end-to-end learning approaches in comparison to classical variational methods are the requirement of expensive retraining for even slightly different problem statements and the lack of provable error bounds during inference. Recent works tackled the first problem by using networks trained for Gaussian image denoising as generic plug-and-play regularizers in energy minimization algorithms. Even though this obtains state-of-the-art results on many tasks, heavy restrictions on the network architecture have to be made if provable convergence of the underlying fixed point iteration is a requirement. More recent work has proposed to train networks to output descent directions with respect to a given energy function with a provable guarantee of convergence to a minimizer of that energy. However, each problem and energy requires the training of a separate network. In this paper we consider the combination of both approaches by projecting the outputs of a plug-and-play denoising network onto the cone of descent directions to a given energy. This way, a single pre-trained network can be used for a wide variety of reconstruction tasks. Our results show improvements compared to classical energy minimization methods while still having provable convergence guarantees.