OCR For Low Resource Languages

This work is an adaptation of CNN+Transformer architecture to training text recognition models for Yorùbá & Igbo.

Architecture

Setup

• Clone the github repository

  $ git clone https://github.com/ToluClassics/LowResourceOCR.git
  

• Create a virtual environment and install dependencies

  $ python2 -m venv venv 
  $ (venv) pip install -r requirements.txt
  

• Download TextRecognition Model from google drive

  • Igbo : Download link
  • Yorùbá : Download link
  • English : Download link

• Run Inference:

  

  python3 inference.py --lang yor 
      --image_path samples/yor_sample.jpg 
      --checkpoint_path run/checkpoint_weights_igbo_trdg.pt
  

  Output

  [INFO] Load pretrained model
  [INFO] Predicted text is: Àwọn Ohun Tó Wà
  

Reference::

• Scene Text Recognition via Transformer, Paper