Vedic Sanskrit Traineddata for 4.0

[Shreeshrii]

• See https://github.com/Shreeshrii/imagessan/tree/master/groundtruthimages for the images used for testing.

san.traineddata in this repo (4.0 alpha version) gives the following accuracy for the above sample of images:

Character/Word Error Rate	%
CER	11.55
WER	8.85
WER (order independent)	7.44

Improved Accuracy is gained after training using
https://github.com/Shreeshrii/tess4training/blob/a09bbe913b25b0e623f1bc267a60b192d8e0ccc6/san.traineddata

Character/Word Error Rate	%
CER	3.73
WER	4.71
WER (order independent)	3.82

The newly traineddata accuracy should improve further , I hope, as the training converges further.

Eval Report at https://shreeshrii.github.io/tess4eval-san/

Update:

Further training actually led to lower accuracy on this sample set.

Character/Word Error Rate	%
CER	6.65
WER	9.43
WER (order independent)	7.94

Stopped Training with
https://github.com/Shreeshrii/tess4training/blob/8dc4f8488e74d4a934168844932c8a526d70c1d9/bihtune.traineddata

[Shreeshrii]

I separately trained for Vedic Sanskrit using text from Rigveda.

The source files are at https://github.com/Shreeshrii/tess4training-vedic

Resulting traineddata files are as follows:

• https://github.com/Shreeshrii/tess4training/blob/master/docs/vedic.traineddata
• https://github.com/Shreeshrii/tess4train/blob/master/docs/rig.traineddata

and accuracy on sample page -
(https://github.com/Shreeshrii/tess4training/blob/master/scanned.tulasi.exp0.tif) when I stopped training was

• https://github.com/Shreeshrii/tess4training/blob/master/docs/BRH-test-vedic_out.html
• https://github.com/Shreeshrii/tess4train/blob/master/docs/BRH-test-rig_out.html

A special thank you to the Travis team for providing the resources for training.

[Shreeshrii]

A user who tested with a sample of 30 pages with the above reported accuracy of 90%.

@theraysmith I hope that the new version of Sanskrit traineddata will be able to OCR both Classical Sanskrit and Vedic Sanskrit. When should we expect the new (4.0.0beta) version of traineddata files?

[Shreeshrii]

Sample of Samaveda Sanskrit text - it uses different set of Vedic accents compared to rigveda sample above. I have not tried training for this.

http://sanskrit.safire.com/image/SamaVeda.gif

http://sanskritweb.net/samaveda/sample.gif

http://vedicreserve.mum.edu/sama_veda/sama_veda.pdf

Update:

Unicode Samaveda text available from http://www.parankusa.org/SamaBrowse.aspx

[ksdmahesh]

how to train
help me by step by step process.
for Vedic sanskrit

[Shreeshrii]

@ksdmahesh What kind of text do you want to train for? You will needs a large text corpus in utf-8 format and unicode fonts that render it correctly.

[ksdmahesh]

ok thank you.
after collection of data.
how to start

[Shreeshrii]

Sorry for delay in reply.

See https://github.com/tesseract-ocr/tesseract/wiki/TrainingTesseract-4.00
for the details about training.

I am doing a test training using portions of online texts of rigveda and yajurveda. (I had deleted the githib repos referenced in earlier messages in this thread). I will share the traineddata file when done.

[bohrbrar]

@Shreeshrii ,
Good work. How close you are in training rigveda and yajurveda models. can you share traineddata files?
Regards
Bohar

[sd-dwivedi]

@Shreeshrii please share the trainedata file for vedic pandulipi sanskrit