Why do I get such bad results?

[Doron Saar]

Hi,

I'm trying to train Tesseract to work with a large library of Hebrew language documents.

They are all in good quality scanning, black and white, and most of them have the same font and character size.

The hebrew alphabet should be relatively very simple for OCR: 27 characters, no Upper/Lower cases, characters seperated from each other and standard punctuation like in English.

Even though, after creating manually about 30 training BOX files and compiling them, I still get very poor results.

(about 70% accuracy).

It does not seem to improve when I add more training data.

What can cause this?

Do I need more training documents?

Is there a minimal characters resolution?

What can I do better?

[Ashish Goel]

If you can elaborate on what kind of failures you are experiencing, people might be able to help.

[Doron Saar]

I just get the same mistakes all the time.

The letter ו is often read as ט 

The letter נ is often read as )

and so on.

When I add more training data files I just get worse results instead of better results.

[Tom Morris]

On Monday, June 6, 2016 at 3:17:29 AM UTC-4, Doron Saar wrote:

I'm trying to train Tesseract to work with a large library of Hebrew language documents.

Why? Did you get unacceptable results with the standard Hebrew language data? 

https://github.com/tesseract-ocr/tessdata/blob/master/heb.traineddata

Tom

[ashish goel]

I had same problem for Swedish language and a temporary workaround helped me. I zoomed (re-scaled) image to 400% and it recognized the letter. (Though it added other problems). Not sure, but it could improve results for you.

Ashish

--
You received this message because you are subscribed to the Google Groups "tesseract-ocr" group.
To unsubscribe from this group and stop receiving emails from it, send an email to tesseract-oc...@googlegroups.com.
To post to this group, send email to tesser...@googlegroups.com.
Visit this group at https://groups.google.com/group/tesseract-ocr.
To view this discussion on the web visit https://groups.google.com/d/msgid/tesseract-ocr/df837034-f9a2-4aa1-889f-4d7614f3d8c8%40googlegroups.com.

For more options, visit https://groups.google.com/d/optout.

[Bojidar Stanchev]

 If there originaly was a training data in the tessdata repository try using this one. Even in tesseract's official repository is stated that retraining is not likely to help you.

 You can try different scales of the image. Are you doing any image processing before giving the images to tesseract. You can familiarize yourself with how tesseract uses a dictionary, maybe there is the problem. Maybe you can process tesseract output - like calculate Levenstein distance from the output to some words from a dictionary.

Bear in mind that 100% recognition is impossible. I would say 80-90% is what you will probably end up with. 90%+ is extreme success in my opinion. Good luck.