Looking promising... »
Had a bit of a brainwave the other night — clustering connected components using size similarity, locality and fit to a line as criteria. The image shows the kind of results I’m getting with a really quick script thrown together late last night.
Blue boxes outline the individual characters that have been recognised, red boxes outline clusters.
It’s sometimes missing some of the characters at the ends of the plates, and there are still some false positives showing up in some images, but I’m fairly confident that a bit of tuning could fix all that.