Jeremy Howard describes this as "Decrappification"[1]. This is one of the easiest deep learning models to train, in my opinion, as you can generate your own dataset easily. You just get good pictures for the target, programmatically make changes that make the image "crappy" for your source, and train until your network can convert from crappy to good. Then you pass it something it has never seen, and whabam, your picture is sharper than before.

[1] - https://www.fast.ai/2019/05/03/decrappify/