Wonder if you could train a neural net to take camera recordings and basically reconstitute the original. For a given setup, the distortions should be pretty consistent.