This should be tagged #dsp.
Deconvolution is a really interesting area, and there’s an immense amount of academic work trying to apply basically every machine-learning and AI trick in the book to the problem. This blog post is a very accessible introduction!
What if you allowed humans to tag pictures with what they thought they were (and maybe train the AI to recognize those objects—yes, I know it’s harder than it sounds) and have the AI learn each differently? Also, what if humans highlighted areas of continuous and discontinuous color change, e.g., buildings and patterns vs. the sky and skin tone? That might help prevent the “ring effect” that occurs.