banglaOCR

Version: 0.0.6     
Authors: Md. Nazmuddoha Ansary 

LOCAL ENVIRONMENT

OS          : Ubuntu 18.04.3 LTS (64-bit) Bionic Beaver        
Memory      : 7.7 GiB  
Processor   : Intel® Core™ i5-8250U CPU @ 1.60GHz × 8    
Graphics    : Intel® UHD Graphics 620 (Kabylake GT2)  
Gnome       : 3.28.2  

Setup

Assuming the libraqm complex layout is working properly, you can skip to python requirements.

• sudo apt-get install libfreetype6-dev libharfbuzz-dev libfribidi-dev gtk-doc-tools
• Install libraqm as described here
• sudo ldconfig (librarqm local repo)

python requirements

• pip requirements: pip3 install -r requirements.txt

Its better to use a virtual environment OR use conda-

• conda: use environment.yml
• install tesseract if you want to use the tesseract-ocr. Make sure to properly setup the bangla data.

Pretraining Dataset

• The overall dataset is available here: https://www.kaggle.com/nazmuddhohaansary/sourcedata
• We only need the bangla folder for this
• The path for the source folder is the data_path used in datagen_pretrain.py and datagen_finetune.py
• The dataset is collected and compiled from vairous sources such as:
  • The bangla grapheme dataset is taken from here.
    • Only the 256 folder under 256_train is kept and renamed as RAW form BengaliAI:Supplementary dataset for BengaliAI Competition
  • The bangla number dataset is taken from here
    • Only the RAW_NUMS folder is kept that contains all the images of the numbers

FineTuning DataSet

• For Modifier Finetuning the following open source dataset was used: https://www.kaggle.com/nazmuddhohaansary/boisebangladata

• This data set is variation of dataset introduced by: Nishatul Majid and Elisa Barney-Smith [DOI](https://doi.org/10.18122/saipl/1/boisestate)

• The folder structre should look as follows:

        source
        ├── bangla
           ├── graphemes.csv
           ├── numbers.csv
           ├── dictionary.csv
           ├── fonts
           ├── graphemes
           └── numbers
        ├── boise_state
            

Execution

• run datagen_pretrain.py
• run datagen_finetune.py