> On Sep 26, 2015, at 5:42 PM, Toby Thain <[email protected]> wrote:
> ...
> Software which "recreates" the typography of a document from OCR does not
> produce an acceptable substitute, I've yet to see a book that wasn't ruined
> by it.
True. But that's not the biggest problem with OCR. The biggest problem is
that even professional grade OCR programs have rather low accuracy. Maybe they
do acceptably well on really high grade scans of very clean new documents, but
on books, typewritten documents, etc., even after you use the "train" feature
you need to spend a long time cleaning up. It may be faster than retyping
things, if you're lucky. Not if you're not; two of us recently retyped 300
pages of line printer listing because that was faster and more accurate than
OCR on that particular printout.
Given that OCR can only do, at best, a just barely acceptable recognition of
the letters of the alphabet, it follows that accurately recognizing the actual
font used will be vastly less accurate. And indeed you can see that clearly.
I wonder if there are OCR programs that can be told to choose among 2 or 3
fonts, as opposed to guess from the entire inventory of the machine. If so,
and if they are sufficiently distinct, then maybe you'd stand a chance.
Especially if it also added heuristics like "never change fonts in mid-word" --
an obvious rule but not one I have seen implemented.
paul
