Add support for Hebrew Language and Alphabet

[johnlockejrr]

https://en.wikipedia.org/wiki/Unicode_and_HTML_for_the_Hebrew_alphabet

Hebrew is like Arabic a RTL Language.

[CLAassistant]

All committers have signed the CLA.

[johnlockejrr]

Seems like the script failed beeing killed

[GreatV]

please fix codestyle

[johnlockejrr]

Fixed

[GreatV]

LGTM

[ephraimm]

Hi @johnlockejrr

I'd like to test the hebrew support, but having trouble getting a compiled version working to test. could you point me in the right direction.

many thanks

[johnlockejrr]

I didn't train yet a model for Hebrew on PaddleOCR, got stuck on work with YOLOv8/PyLaia and kraken for Hebrew, and with Hebrew I mean classical Hebrew, medieval not Ivrit

[luotao1]

@johnlockejrr Thanks for your contribution! You will receive a beautiful PaddlePaddle gift. Please provide your mailing address by filling out the following questionnaire before October 18th.

Looking forward to the future, we will walk further together in the world of open source!
Click Here ：https://paddle.wjx.cn/vm/h4On9gJ.aspx#

[ephraimm]

Is there anyway I can help. What are the next steps?

[johnlockejrr]

@ephraimm, sure, I'm right now collecting as much ground-truth/datasets of very good quality for Hebrew. In my case, I mostly work with ancient/medieval Hebrew (biblical, rabbinical, responsa etc. stuff) but Ivrit also. We can collaborate if you want.

[ephraimm]

@johnlockejrr we're after the same thing.
My end game is to digitise printed (chassidic) texts that usually have complex section breaks and columns. I have had some success combining tesseract with google vision, but this is not ideal and would like to find an open source solution.

tesseract is good at recognising blocks, but not good at distinguishing between actual text and subscript references to footnotes and punctuation. google on the other hand is great at accurately recognising the characters but doesn't recognise column breaks etc.

My current hypothesis is if I give tesseract a good quality sample of text from google and fine tune the existing Hebrew model we'll get the results.

What makes you think paddle will work? Have you tried fine tuning tesseract?

[johnlockejrr]

Then take a look at eScriptorium that uses kraken but also can use tesseract (that is not actively developped anymore).
Printed texts are piece of cake for kraken, though you have to train a good segmentation (text and layers detection) model. I mostly work with manuscripts, HTR is a little more difficult than printed texts.

Drop me an email [at] gmail.com

[ephraimm]

@johnlockejrr
sent to gee male. you can do likewise

[luotao1]

@johnlockejrr Could you provide your email? We need this information when mailing gifts. Please send your email to ext_paddle_oss@baidu.com or here. Thanks very much!

[johnlockejrr]

@luotao1 is: johnlockejrr [at] gmail.com