OCR work at GiellaLT

This page will at some point document our OCR work.

We have been experimenting with OCR in 2016 and earlier (cf. the meeting memo from 2016 below). With recent advances in OCR techniques we will have to start this work again, with new programs. This page sketches how.

Experimenting with Tesseract

Fetching the program

The open source program Tesseract can be fetched from Github:

git clone git@github.com:tesseract-ocr/tesseract.git
git clone git@github.com:tesseract-ocr/tessdata.git
...

Development

Tesseract comes with a set of languages (see tessdata). Most GiellaLT languages are not included, though. TODO: Document how to add them.

OCR reading

A pdf document as a picture should be

1. split into one pdf per page
2. converted to png
3. run through Tesseract with the relevant language(s) as setting

One pdf per page

In Preview, set the document in Thumbs view and drag one page at a time to the desktop. TODO: Find a way to do this on the command line.

converted to png

Let us say the document contained 8 pages, after the split named 1.pdf, 2.pdf, … Then do the following:

for i in 1 2 3 4 5 6 7 8 9 10 ; do sips -s format png $i.pdf --out $i.png ; done

run through Tesseract

Let us say the document contains Norwegian and Finnish. Standing in tesseract-ocr/tesseract, run the 8 pdf files through tesseract:

for i in 1 2 3 4 5 6 7 8 ; do tesseract --tessdata-dir ../tessdata/ $i.png $i.txt -l fin+nor ; done

The resulting files may then be collected into one text file.

Meetings

2016

• 15.12