Hi, all. Jonathan has been very kind in promoting Nick's work and mine here. The following is very much 'Inside Baseball'. If you aren't interested in the details of Greek OCR, just be aware that this field is shifting rapidly and improvements are being made monthly. OCR is never perfect, but we are getting many results on very difficult texts that are way better than even a year ago. See, for instance, Migne vol. 45:
The scan is pretty atrocious, but we get some pretty good output.
If you want the details, then a quick word about terminology. 'Lace' is the name I'm using for the website and related code which allows us to view facing page images and html OCR output. The code for this is in github https://github.com/brobertson/Lace
, and our site is at http://heml.mta.ca/lace
. It currently has 780 books in it, all generated with resources from Compute Canada.
The output of any OCR engine that generates html can be viewed in Lace. Tesseract is one such engine. Nick, now working with the Perseus group in Boston, is a whiz at Tesseract, initially training it as a side
project to a larger effort in the UK!
Gamera is another engine. Federico Boschetti and I started working with it a couple of years ago and last year built an entire HPC stack around eeking out the best OCR from Gamera and the OCR that archive.org already provides. This stack, called Rigaudon, includes automatic spellcheck and OCR output combining. It was with Gamera/Rigaudon that the first 500-600 books in the Lace repository were generated. Running Rigaudon at home requires a multi-core machine and some pretty sophisticated software.
Another promising engine is Ocropus 0.7. Its basic recognition is very impressive, it can be trained fairly simply, and it is able to deal with a variety of letter thicknesses, etc. One downside is it requires much training data; however, we can use the data from Gamera/Rigaudon to train it! I've been focusing on this since about January, making classifiers for Porson and Teubner fonts, and moving on from that. To get the very best results, though, we need to do all the additional things that Rigaudon provided to Gamera. The complete Ocropus stack for polytonic Greek is called Ciaconna, and like the rest is available in github.
Ultimately, we will use all three engines in harmony with each other. Federico wrote a paper on this in, I think 2008, so we know how to do it. But right now, it makes sense to work up each engine separately. We are focusing on these because they are free and open-source.