Multilingual OCR Development Plan #12734

Traditional Mongolian

I would love to work on "Bangla"

I very happy if you do that with Vietnamese

How about Arabic? That would be great.

I've find out that PADDLE OCR algorithm cannot recognize some special characters (such as comma, semicolon, or dot...) when the language is english. Is there any possible way that i can fix this problem

I would like to contribute to add the Burmese language. Is it only needed to submit two text files - dict & corpus? How further process do we need to provide?

Adding "Bangla" will be grate for the people in south Asia

Adding "Traditional Chinese (zh-TW)" would be great support.

Do you have preTrained Russian recognition model?

Hi adding " Tamil" language will be very grateful.

Tamil_dict.txt
Tamil_corpus.txt

Need more help plz refer this issue:
JaidedAI/EasyOCR#39

I can help with polish language.

@GmGniap Hello, Can you provide the corpus file of Burmese Language？

@shahidul56 Hello, Can you provide the corpus file of Bangla Languag？

All models updated in 2021.1.21 cannot be downloaded with following Error：
{ code: "NoSuchKey", message: "The specified key does not exist.", requestId: "aa1bfeff-f572-40aa-8935-6129b1533ed1" }

All models updated in 2021.1.21 cannot be downloaded with following Error：
{ code: "NoSuchKey", message: "The specified key does not exist.", requestId: "aa1bfeff-f572-40aa-8935-6129b1533ed1" }

Sorry for the invalid links and all of them have been revised now, you can try again.

Truong

I'm trying with my private data, but the result very poor

Sorry for my stupid question, I am novice at DL: What difference between Inference model and trained model?

I created a PR for Bangla

Does this list contain the latest models? If i want to fine tune for example german model do i use this link from this page to download the pretrained model? If so what yml file should i use? How do i know what is the architecture of these models?

Please add Tajik Language
tajik_corpus.txt
tajik_dict.txt

I want to work on Kurdish Center language

I sent a PR for Bangla support #13373
btw do you not have the corpus folder anymore?

pls add Turkish languange.Thank you.

How do i use these models?

I have trained a model on persian using synthetic data the results are close but don't have access to the data the PaddleOCR had so it has weak characters w. ref back to the fa_dict and arabic_dict.txt. the results are available on kaggle under name farboodi/persian-OCR

bengali_corpus.txt
bengali_dict.txt
I would like to contribute to bangla integration. please help us to do so as million people waiting for this. I have attached both files. Please let me know what else to provide.

Hebrew is missing from this list — here is a charset and a working set of models

Hebrew doesn't appear anywhere in this thread: not in the supported list, not in
the requests. Arabic is covered, so the RTL machinery is already there, but
Hebrew has no PP-OCR recognition model at all (nor does MMOCR or EasyOCR).

We have been training Hebrew on PaddleOCR for a while and have released the
result, so rather than filing a request we would rather contribute what we
already have.

1. Character dictionary. A 120-character Hebrew set covering the alphabet
including final forms, Latin, digits and the punctuation that actually shows up
in Hebrew documents (including geresh/gershayim, which are routinely confused
with ASCII quotes and are a real source of error):
https://huggingface.co/Rivok/paddleocr-hebrew/blob/main/charset_v2f.txt

2. Working models, Apache-2.0. Finetuned from PaddleOCR v3.7.0 — a Hebrew
mobile DBNet detector, plus recognizers built on PP-OCRv5, PP-OCRv6 and SVTRv2.
Released as ONNX: https://huggingface.co/Rivok/paddleocr-hebrew
Code and full writeup: https://github.com/RivoksLab/paddleocr-hebrew

3. Two findings that may affect other RTL languages here, Arabic included.

• A monotonic CTC head deletes Latin/digit runs embedded inside RTL text —
  blank-collapses them rather than misreading them. Isolated Latin reads at
  1.20% CER and isolated digits at 1.41%, but Hebrew-with-embedded-Latin is
  12.63%. Swapping to an attention head on the same backbone takes that to
  2.33%. If Arabic PP-OCR users are seeing Latin runs vanish inside Arabic
  lines, this is likely why.
• Long lines: line-level recognition warm-started from a word-specialised
  base cliffs at roughly its training crop length. SVTRv2 reads long lines
  single-pass and removed the plateau entirely for us.

Happy to help get Hebrew into the official multilingual set. Numbers above are
in-house on our own test sets — vendor-reported, not independently verified.