rakeshvar / chamanti_ocr

Chamanti OCR చామంతి

Mission

This project aims to build an ambitious state-of-the-art OCR framework, that should work on any language. It will not rely on segmentation algorithms (at the glyph level), making it ideal for highly agglutinative scripts like Arabic, Devanagari etc. We will be starting with Telugu however. We use the technology of Convolutional Recurrent Neural Networks from Keras in TensorFlow 2.0. CRNN with CTC (Connectionist Temporal Classification) loss function is the main work-horse.

Dependencies

1. tensorflow
2. Lekhaka - My 'scribing' package for generating complex text, including Indian languages like Telugu, on the fly

Setup

1. Install TensorFlow
2. Download Lekhaka and place in a parallel dicrectory

Files

1. model_builder.py The TensorFlow CRNN model with CTC loss
2. train.py Main file to run
3. utils.py, post_process.py Utilities to print images and Probabilities to terminal, etc.

Training the CRNN

You can now train a CRNN to read Telugu text!

python3 train.py spec 1
python3 train.py banti banti_trained_instance.pkl
python3 train.py chamanti chamanti_trained_instance.pkl