That's what I was thinking. Manually reviewing 20,000 papers is simply an astonishing feat. Heck, even searching a single paper for anything that looks like anything else in the same paper (possibly rotated, scaled, or alpha blended) might be beyond me. But it's the kind of task a computer vision algorithm should be able to do very well at. In fact, I Googled it and found (what else) some papers about it, e.g.:
https://arxiv.org/abs/1407.6879