[Suggestion] "Training light" - Learning by doing - re-feeding a corrected text file to retrain tesseract

[Wikinaut]

(probably something for #1423)

When I have a manually corrected text version (txt output file), my deep wish is to have an easy way to retrain tesseract, based upon the 1. first tesseract txt output and the related 2. manually corrected txt file.

I remember that I asked this question already some years ago. But now situation has changed (more and other developers, new algorithms).

Do you like the idea? Is it possible?

[stweil]

The idea is good, it is possible, but today the way is not easy. So there remains work to be done to make it easy.

[Wikinaut]

@stweil Can I help with a donation?

[stweil]

See http://gepris.dfg.de/gepris/projekt/394264782 (German).

[Wikinaut]

Ich verstehe. Developed a server some years ago (open-source replacement for Finereader appliance) with a scheduler and so on. Users could upload multi-page pdf documents with a web formular upload (no limit, we could ocr >1.000 page-pdfs), and users received an e-mail and a link when their "ocr jobs" were ready.

[Wikinaut]

This was when I discovered that tesseract (then) internally re-coded documents with a lossy compression, which is a "NOGO" for CG and something with sharp edges: like fonts, the thing we want to ocr....

[amitdo]

As @stweil already said, this feature was implemented years ago (fine tuning a lstm model).

We should not keep this 'issue' open for the 'make it easy' part. PRs to this repo or to the tesstrain repo are welcomed.