I would think about feeding rendered unicode characters to train a visual algorithm about which character may match, then "on the go" do a bitmap comparison to

I would think about feeding rendered unicode characters to train a visual algorithm about which character may match, then "on the go" do a bitmap comparison to select the best one for the given case. The first part could be something like shapecatcher.com where you select the top N plausible candidates, then on the final part you select the best one for the case at hand. (well at least that's how I would try to implement that first) Anyway yes, it would be great!!