Adding New Fonts to Tesseract 3 OCR Engine

2014-10-06T00:00:00-07:00

Status

I’m attempting to keep this up to date as Tesseract changes. If you have corrections, please send them directly using the contact page.

I’ve turned off commenting on this article because it was just a bunch of people asking for help and never getting any. If you need help with these instructions, go to Stack Overflow and ask there. If you send me a link to a question on such a site, I’m much more likely to respond positively.

Introduction

Tesseract is a great and powerful OCR engine, but their instructions for adding a new font are incredibly long and complicated. At CourtListener we have to handle several unusual blackletter fonts, so we had to go through this process a few times. Below I’ve explained the process so others may more easily add fonts to their system.

The process has a few major steps:

Create training documents
Teach Tesseract about the documents

Create training documents

To create training documents, open up MS Word or LibreOffice, paste in the contents of the attached file named ‘standard-training-text.txt’. This file contains the training text that is used by Tesseract for the included fonts.

Set your line spacing to at least 1.5, and space out the letters by about 1pt. using character spacing. I’ve attached a sample doc too, if that helps. Set the text to the font you want to use, and save it as font-name.doc.

Save the document as a PDF (call it [lang].font-name.exp0.pdf, with lang being an ISO-639 three letter abbreviation for your language), and then use the following command to convert it to a 300dpi tiff (requires imagemagick):

convert -density 300 -depth 4 lang.font-name.exp0.pdf lang.font-name.exp0.tif

You’ll now have a good training image called lang.font-name.exp0.tif. If you’re adding multiple fonts, or bold, italic or underline, repeat this process multiple times, creating one doc → pdf → tiff per font variation.

Train Tesseract

The next step is to run tesseract over the image(s) we just created, and to see how well it can do with the new font. After it’s taken its best shot, we then give it corrections. It’ll provide us with a box file, which is just a file containing x,y coordinates of each letter it found along with what letter it thinks it is. So let’s see what it can do:

tesseract lang.font-name.exp0.tiff lang.font-name.exp0 batch.nochop makebox

You’ll now have a file called font-name.exp0.box, and you’ll need to open it in a box-file editor. There are a bunch of these on the Tesseract wiki. The one that works for me (on Ubuntu) is moshpytt, though it doesn’t support multi-page tiffs. If you need to use a multi-page tiff, see the issue on the topic for tips. Once you’ve opened it, go through every letter, and make sure it was detected correctly. If a letter was skipped, add it as a row to the box file. Similarly, if two letters were detected as one, break them up into two lines.

When that’s done, you feed the box file back into tesseract:

tesseract eng.font-name.exp0.tif eng.font-name.box nobatch box.train .stderr

Next, you need to detect the Character set used in all your box files:

unicharset_extractor *.box

When that’s complete, you need to create a font_properties file. It should list every font you’re training, one per line, and identify whether it has the following characteristics: <fontname> <italic> <bold> <fixed> <serif> <fraktur>

So, for example, if you use the standard training data, you might end up with a file like this:

eng.arial.box 0 0 0 0 0
eng.arialbd.box 0 1 0 0 0
eng.arialbi.box 1 1 0 0 0
eng.ariali.box 1 0 0 0 0
eng.b018012l.box 0 0 0 1 0
eng.b018015l.box 0 1 0 1 0
eng.b018032l.box 1 0 0 1 0
eng.b018035l.box 1 1 0 1 0
eng.c059013l.box 0 0 0 1 0
eng.c059016l.box 0 1 0 1 0
eng.c059033l.box 1 0 0 1 0
eng.c059036l.box 1 1 0 1 0
eng.cour.box 0 0 1 1 0
eng.courbd.box 0 1 1 1 0
eng.courbi.box 1 1 1 1 0
eng.couri.box 1 0 1 1 0
eng.georgia.box 0 0 0 1 0
eng.georgiab.box 0 1 0 1 0
eng.georgiai.box 1 0 0 1 0
eng.georgiaz.box 1 1 0 1 0
eng.lincoln.box 0 0 0 0 1
eng.old-english.box 0 0 0 0 1
eng.times.box 0 0 0 1 0
eng.timesbd.box 0 1 0 1 0
eng.timesbi.box 1 1 0 1 0
eng.timesi.box 1 0 0 1 0
eng.trebuc.box 0 0 0 1 0
eng.trebucbd.box 0 1 0 1 0
eng.trebucbi.box 1 1 0 1 0
eng.trebucit.box 1 0 0 1 0
eng.verdana.box 0 0 0 0 0
eng.verdanab.box 0 1 0 0 0
eng.verdanai.box 1 0 0 0 0
eng.verdanaz.box 1 1 0 0 0

Note that this is the standard font_properties file that should be supplied with Tesseract and I’ve added the two bold rows for the blackletter fonts I’m training. You can also see which fonts are included out of the box.

We’re getting near the end. Next, create the clustering data:

mftraining -F font_properties -U unicharset -O lang.unicharset *.tr 
cntraining *.tr

If you want, you can create a wordlist or a unicharambigs file. If you don’t plan on doing that, the last step is to combine the various files we’ve created.

To do that, rename each of the language files (normproto, Microfeat, inttemp, pffmtable) to have your lang prefix, and run (mind the dot at the end):

combine_tessdata lang.

This will create all the data files you need, and you just need to move them to the correct place on your OS. On Ubuntu, I was able to move them to;

sudo mv eng.traineddata /usr/local/share/tessdata/

And that, good friend, is it. Worst process for a human, ever.

Enclosures

The Winning Font in Court Opinions

2012-01-27T22:15:58-08:00

At CourtListener, we’re developing a new system to convert scanned court documents to text. As part of our development we’ve analyzed more than 1,000 court opinions to determine what fonts courts are using.

Now that we have this information, our next step is to create training data for our OCR system so that it specializes in these fonts, but for now we’ve attached a spreadsheet with our findings, and a script that can be used by others to extract font metadata from PDFs.

Unsurprisingly, the top font — drumroll please — is Times New Roman.

Font	Regular	Bold	Italic	Bold Italic	Total
Times	1454	953	867	47	3321
Courier	369	333	209	131	1042
Arial	364	39	11	41	455
Symbol	212	0	0	0	212
Helvetica	24	161	2	2	189
Century Schoolbook	58	54	52	9	173
Garamond	44	42	41	0	127
Palatino Linotype	36	24	24	1	85
Old English	42	0	0	0	42
Lincoln	27	0	0	0	27