> a) people aren't willing to put up with false positives (i.e., adblockers hiding the meme cat photos you were looking for)

You can estimate the actual ratio of false positives by examining existing automation solutions based on image recognition. As a very rudimentary example, I have successfully used Sikuli [1] for various tasks including a program running "binary blobs". Unless you mix contents and ads altogether (and I personally think that it is fine as long as the ad is properly marked! It is at least a kind of contents, just that what I don't want generally though.) you can't ramp the false positive ratio up beyond a certain level.

[1] http://www.sikuli.org/