Here's a good one.

I'm looking for a reference corpus of scanned data for OCR testing. It has to be free, it has to be scanned at at least 200dpi and preferable 300dpi, and it has to be available to non-academic researchers like me.

I've found some of the free newspapers work moderately well, but they all seem to be scanned at too low a resolution to be able to get sane results from OCR - though they can usually be read ok.

I need about a million words, preferably three or four million... that's about five King James Bible's worth.

Any thoughts?

You could possibly try Car Workshop repair Manuals.
Might just mean registering at Car Club first. I do remember downloading gigabytes of scanned repair manuals for one of my old cars years ago.
I'll try to find the site again - if it still exists.

Well what about our own (fccuk) scanned copy of the Coupe workshop manual? Might as well produce something useful at the end of it!

Good thoughts but text is preferable to images.

http://www.allonlinebible.com/read/ x 5

Needs to be scanned text Pete.

There is text in the Coupe workshop manual, pictures too, but still a lot of text

But not three million words.

The problem is that the good OCR products are too good: if I just feed it a clean image of a book I get *no* errors in a hundred thousand words. As I'm trying to do so tests that require errors to be there, that's useful but a little pointless...

I've been playing with a complex but repeatable pre-process method that involves turning the text into a pdf, extracting the pages from the pdf with a low resolution and saving as jpegs, and then doing the OCR. That gives me introduced noise from the scaling and the jpeg artifacts which introduces some errors to the OCR.

I guess it will have to do.

I found the ultimate test for OCR was scanning in pages from a book my father wrote on an electronic typewriter 7 years ago. The ink ribbon was past its best and the top of the characters were only half or so the density of the bottom half. I resorted to retyping almost everything because the OCR programs really couldn't make much out of it. I probably still have the scans. Can you do something with that?

It would be interesting to find out, though I couldn't use it as a reference for a published paper. You must have an email for me? Current OCR programs are very very good...

Was your father's book written in English, Klaas? Neil, I thought your algorithm only worked on English text?

Ah, yes... forgot that. Though it would still be interesting to see what happens. I theorise it would work with Latin languages, less so with French (yes, yes, I know) and German. Though there are enough French and German influences in British spelling that Low German might work...