recognising roman with sanskrit diacritics

[yajva]

I have tried Google OCR for recognizing Sanskrit text in Roman with diacritics (IAST). It recognizes above macron but not dots below also joining grave and accent. Is there any traineddata available for tesseract that can do this with good accuracy ? Attached a sample page that I am interested in.

[Shree Devi Kumar]

I had done a training for sanskrit for both devanagari and IAST but it does not include cedilla for Sh 

I will add it and let you know.

On Wed 20 Jun, 2018, 1:17 AM yajva, <nsvnar...@gmail.com> wrote:

I have tried Google OCR for recognizing Sanskrit text in Roman with diacritics (IAST). It recognizes above macron but not dots below also joining grave and accent. Is there any traineddata available for tesseract that can do this with good accuracy ? Attached a sample page that I am interested in.

--
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/aef0797b-8df3-4db7-9a3b-02f62d2e5a28%40googlegroups.com.
For more options, visit https://groups.google.com/d/optout.

[Shree Devi Kumar]

I am attaching the OCRed text. Please correct it so that  I can use as groundtruth for further training and testing.

--

____________________________________________________________
भजन - कीर्तन - आरती @ http://bhajans.ramparivar.com

[yajva]

done

[yajva]

one more correction.

[Shree Devi Kumar]

Please try with iast.traineddata model for tesseract.4.0.0-beta posted at https://github.com/Shreeshrii/tessdata_sanskrit

To view this discussion on the web visit https://groups.google.com/d/msgid/tesseract-ocr/a7bdf637-7f17-4eb3-8fa8-297018633bfa%40googlegroups.com.

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

[Shree Devi Kumar]

Sorry, there seems to be some regression in the file posted on github. I will upload again later.

[Shree Devi Kumar]

Please test with traineddata file from https://github.com/Shreeshrii/tessdata_sanskrit/tree/master/iast-plus1

Need to check that is it not overfitted.

Please share a couple more images which I can use for testing.

On Thu, Jun 21, 2018 at 11:38 PM yajva <nsvnar...@gmail.com> wrote:

To view this discussion on the web visit https://groups.google.com/d/msgid/tesseract-ocr/a7bdf637-7f17-4eb3-8fa8-297018633bfa%40googlegroups.com.

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

[yajva]

Sorry for the delay, my system was down.

I am getting "Page not Found" for the link given. Can you pl re-check?

Here's the doc I am trying to OCR

[Shree Devi Kumar]

Traineddata file is attached for use with tesseract4.0.0-beta.

How did you create the test png from the pdf? I am not getting as good quality, tried various settings with irfanview.

To view this discussion on the web visit https://groups.google.com/d/msgid/tesseract-ocr/81b2b741-471c-45a5-adef-48330d960d62%40googlegroups.com.

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

[yajva]

The doc is diff ver of the same text. Here's the doc used for the first. png. This is slightly darker, but the one sent earlier is cleaner. Let me know which is more amenable for OCRing. I use PDF Shaper to extract images and convert to png using xnview.

[Shree Devi Kumar]

I had used ghostview to convert PDF to tif or png.

You can ocr PDF directly with gimagereader using the traineddata file I sent.

See links for new windows binaries in msg below.

At last, here are some fresh builds:

https://smani.fedorapeople.org/tmp/gImageReader_3.2.99_qt5_i686_tesseract4.git87635c1.exe
https://smani.fedorapeople.org/tmp/gImageReader_3.2.99_qt5_x86_64_tesseract4.git87635c1.exe

I'd be also interested in testing of the tessdata manager, which should now also properly handle script tessdatas

To view this discussion on the web visit https://groups.google.com/d/msgid/tesseract-ocr/ed565236-146d-4902-b3e2-13445939a2f4%40googlegroups.com.

[yajva]

Checked with both light & dark pdfs. The results are very good. Thanks.

A few concerns. E is consistently missed in both. J is missed consistently in darker image but recognized as T in dark image. ṝ is recognized as ṛ consistently. Can these be addressed ?
I am using tesseract 4 alpha windows build from command line.

Are the dev files in repos ?

[Shree Devi Kumar]

ok. I will take a look.

To view this discussion on the web visit https://groups.google.com/d/msgid/tesseract-ocr/f942f9b9-a767-4d9e-9de7-0855179db9b5%40googlegroups.com.

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

[Shree Devi Kumar]

I have uploaded a new version of traineddata file at 

https://github.com/Shreeshrii/tessdata_shreetest/blob/master/iast-layer-18003.traineddata

Attached is the OCRed output for pages 13-24 of dark pdf with it.

I am still training a different variation.

[yajva]

Many thanks. Downloaded and using.
Will wait for next ver.

[yajva]

shree
namaste

I am trying to OCR the attached image. Getting not so good results. Even for text which is apparently clear. Eg. in the first line, B is recognized as H, under dot for 't' in 'most' 4th line etc. The image has warping but
still best/Latin and Google OCR produce better results. Is it possible to add diacritics to Latin? Can you help in any way?

regards
Venkatesh

[Shree Devi Kumar]

What about ocr with 

eng+iast

To view this discussion on the web visit https://groups.google.com/d/msgid/tesseract-ocr/1692f4a3-f536-4e57-b666-5f0c6155514e%40googlegroups.com.

[yajva]

eng+iast-plus-3600 => no diacritics at all
Latin+iast-plus-3600 => only macrons none other

[Shree Devi Kumar]

Thank you for your feedback of eng+

I will try training for this and get back.

To view this discussion on the web visit https://groups.google.com/d/msgid/tesseract-ocr/d2439fb8-2fa7-4988-8b5f-ea23f0fbf4f4%40googlegroups.com.

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