Adding Karakalpak Language to Tesseract OCR

---

Overview

This guide details the process of adding support for the Karakalpak language to the Tesseract OCR engine. It involves setting up the environment, preparing training data, training the model, and generating the Karakalpak .traineddata file.

STEP-1: Info

Dataset Requirements for Tesseract OCR Training

This document outlines the necessary components and formats for creating a dataset for Tesseract OCR training, specifically tailored for the Karakalpak language.

Training Text (.training_text)

Requirements

• Large, diverse set of sentences or phrases in Karakalpak.
• Covers a wide range of vocabulary and sentence structures.
• Represents actual language usage, including common phrases and expressions.

Example

Meniń atım Davron.
Búgın hawa ayaz.
Adamlar paydalanatuǵın gezlemler tiykarınan tábiyǵıy hám ximiyalıq
talshıqlardan alınadı.
Men Qaraqalpaqstanda jasayman

Wordlist (.wordlist)

Requirements

• List of words in Karakalpak, one per line.
• Includes a comprehensive range of commonly used words.
• Special focus on words with unique Karakalpak characters.

Example

men
sen
ol
biz
sizler
olar
kitap
qalam
mektep
adamlar

Unicharset (.unicharset)

Requirements

• Lists all unique characters found in the training text.
• Each line represents a character and its properties.

Example

a 0 Common 0
b 1 Common 0
Á 2 Latin 1
...

Unicharambigs (.unicharambigs)

Requirements

• Specifies ambiguities in character recognition.
• Helps Tesseract understand potentially confusing character sequences.

Example

1    sh    s h    0

Config Files

Requirements

• Contain settings and parameters for Tesseract training.
• Include language-specific configurations.
• Tailored to the Karakalpak language.

Example

tessedit_char_whitelist 0123456789ABCDEFGHIJKLMNOPQRSTUVWXYZÁǴÍŃÓÚShCh

Box/Tiff Files for Training

Requirements

• Box files map text regions in an image.
• Tiff files are corresponding image files.
• Both should be accurately aligned.

Example

• image.tif - Image file.
• image.box - Corresponding box file.

Font Properties File (.font_properties)

Requirements

• Information about fonts used in training.
• Includes font name, styles (italic, bold, etc.).

Example

Arial 0 0 0 0 0
TimesNewRoman 0 0 0 1 0

Adjust these examples as needed to fit the specific characteristics of the Karakalpak language.

STEP-2: Start

Prerequisites

• Google Colab or a similar environment.
• Basic Python programming and command-line knowledge.
• Tesseract OCR and related dependencies.

Setup Instructions

1. Environment Setup

Mount your Google Drive and navigate to your Tesseract directory:

from google.colab import drive
drive.mount('/content/drive')

Navigate to Tesseract directory

git clone https://github.com/qaraqalpaq/preparing_ocr_dataset.git tesseract
chmod 777 -R /content/drive/MyDrive/tesseract
cd /content/drive/MyDrive/tesseract

2. Install Required Libraries

sudo apt install tesseract-ocr
sudo apt-get install tesseract-ocr libtesseract-dev libleptonica-dev pkg-config
sudo apt-get install libicu-dev libpango1.0-dev libcairo2-dev
pip install pytesseract

3. Preparing Training Data

Check the training text file format:

def check_file_requirements(file_path):
    # Implementation here

Generate training images:

text2image --text=./langdata/kaa/kaa.training_text --outputbase=./box-training/kaa --font='Times New Roman' --fonts_dir=./fonts

4. Training the Model

Begin the training process:

tesseract ./kaa.tif ./lstmf/kaa_model_output --psm 6 lstm.train

Add ready kaa_model_output.lstmf file to training_files.txt

ls ./lstmf/*.lstmf > ./lstmf/training_files.txt

5. Additional Training Steps

• Extract unicharset

!unicharset_extractor --output_unicharset ./langdata/kaa/kaa.unicharset ./box-training/kaa.box

• Create a wordlist from kaa.traintext

import re

def create_wordlist_from_train_text(input_path, output_path):
    with open(input_path, 'r', encoding='utf-8') as file:
        text = file.read()

    # Apply your regex transformations here
    # Add your transformations as shown in the previous example

    # Split text into words and remove duplicates
    words = set(re.findall(r'\b[a-zA-ZÁáÚúÓóÍıǴǵÚúŃńÓó]+\b', text))

    # Write the unique words to the output file
    with open(output_path, 'w', encoding='utf-8') as file:
        for word in sorted(words):
            file.write(word + '\n')

# Example usage
train_text_path = './langdata/kaa/kaa.training_text'
wordlist_output_path = './langdata/kaa/kaa.wordlist'
create_wordlist_from_train_text(train_text_path, wordlist_output_path)

Combine model components.

combine_lang_model \
  --input_unicharset ./langdata/kaa/kaa.unicharset \
  --script_dir ./langdata \
  --words ./langdata/kaa/kaa.wordlist \
  --numbers ./langdata/kaa/kaa.numbers \
  --puncs ./langdata/kaa/kaa.punc \
  --output_dir ./output-combine/ \
  --lang kaa

• Modify configurations if needed.

combine_tessdata -o ./output-combine/kaa/kaa.traineddata \
  ./langdata/kaa/kaa.unicharset

• Extract .lstm from english model or old trained model

  combine_tessdata -e ./eng-model/eng.traineddata ./eng-model/extract_eng/eng.lstm
  
  

  or

  combine_tessdata -e ./finish/kaa.traineddata ./eng-model/extract_kaa/kaa.lstm
  

6. Fine-tuning and Evaluating the Model

• Starting LSTM Training for a New Language using an English Model

  lstmtraining \
  --continue_from ./eng-model/extract-eng/eng.lstm \
  --model_output ./trained_models/kaa \
  --traineddata ./output-combine/kaa/kaa.traineddata \
  --train_listfile ./lstmf/training_files.txt \
  --max_iterations 1000

• Fine-tuning or Retraining an Existing LSTM Model for a New Language

  lstmtraining \
      --continue_from ./eng-model/extract-kaa/kaa.lstm \
      --old_traineddata ./finish/kaa.traineddata \
      --traineddata ./output-combine/kaa/kaa.traineddata \
      --train_listfile ./lstmf/training_files.txt \
      --model_output ./trained_models/kaa \
      --max_iterations 1000

or

 lstmtraining \
 --continue_from ./trained_models/kaa.checkpoint \
 --model_output ./trained_models/kaa \
 --traineddata ./output-combine/kaa/kaa.traineddata \
 --train_listfile ./lstmf/training_files.txt \
 --max_iterations 1000

7. Additional Scripts and Utilities

• Sentence generator from wordlist.
• Merge unicharsets.
• Parse HTML content.
• Sort unicharset.

8. Finalizing the Trained Model

Convert the checkpoint files into a .traineddata file:

lstmtraining --stop_training \
              --continue_from ./trained_models/kaa.checkpoint \
              --traineddata ./output-combine/kaa/kaa.traineddata \
              --model_output ./finish/kaa.traineddata

Conclusion

You should now have a functional .traineddata file for Karakalpak language, ready for use with Tesseract OCR.

Using ready model

export TESSDATA_PREFIX=./finish/
tesseract image.jpg output -l kaa

or

![Open Instruction TesseractRestAPI]

Clone rest api service for converting image/pdf to text

git clone https://github.com/qaraqalpaq/TesseractOCR-API.git tesseract-api
cd tesseract-api

Install docker-compose and run project

• For more information about Docker Compose, refer to the official Docker Compose documentation.

sudo apt update
sudo apt install docker-compose
docker-compose up -d

Request :

POST http://localhost:3002/upload
Form-data:
  file: [file content]
  language: 'kaa'

Response:

{
  "message": "File received",
  "task_id": "task-uuid",
  "file_uuid": "file-uuid"
}

Downloading Processed Text File

Request:

GET /download/file-uuid.txt

Notes

• Ensure accuracy in paths and filenames.
• Regularly save progress.
• Use representative training data for best results.