Seems like there are maybe 3+ tangential problems here. Noise is one. Then for your 'simple' list, most of those are the result of direct non-linear transformations, I would imagine an old-school mix of convolution and trad neural nets could come up with their inverses fairly efficiently. The 'complex' list - hmm, the word Deep springs to mind ...