Learned Gradient Descent Performance in Bicubic Super-Resolution Task

In most cases, being an ill-posed problem, image restoration opts to restore a high-quality image from a low-quality one, assuming that some degradation model produced given low-quality input. A lot of restoration methods were proposed for the case when linear degradation operator and i.i.d. Gaussian likelihood is assumed. However, such methods are known not to generalize well. They show a sub-par performance on real data, for which the actual degradation model is neither linear nor even exactly known. The state-of-the-art machine learning allows for overcoming this issue and learn a restoration model to the real data. The main drawback of such approaches is overfitting since to learn an inverse mapping between the low-quality and high-quality samples, they rely entirely on data. They do not utilize limited but existing knowledge of how degradation was performed. In this paper, we study learned gradient descent based image restoration and synthesis. Both linear and non-linear known restoration problems are considered, and research on how a known degradation model may be incorporated in a learned gradient-based restoration procedure is provided. Our results demonstrate that explicit usage of the degradation model and its learned linear and non-linear approximations boost restoration quality compared to a baseline without this feature.