Scene Text-Spotting based on PSEnet+CRNN

Pytorch implementation of an end to end Text-Spotter with a PSEnet text detector and CRNN text recognizer. We plan to grow this repository into an open research platform for multi-lingual text detection and recognition from natural scene images, targeted towards low-resource languages.

Requirements

• Python 3.6.5
• Pytorch 1.2
• pyclipper
• Polygon 3.0.8
• OpenCV 3.4.1

Demo

• Download the trained CRNN and PSEnet models from the links provided below.
• Copy paths of the models and paste them in params.py
• run end-end.py

python end-end.py --img [path to image] --e2e_config_name [end to end config name]

Pre-trained Models

Both PSEnet and CRNN pre-trained models can be found here: gdrive

• the PSEnet model is a multi-lingual text detector, trained on MLT 2019. Works quite well!
• the CRNN recognizes Hindi, Bangla, Malayalam, Kanada, Tamil, Telugu, Odia, Sanskrit, Marathi!

Download the models in models/ directory and modify params.py if required.

Training instructions

• To train your own detection model refer to this file.
• To train your own recognition model refer to this file.

Samples

Contributors

• Azhar Shaikh, PES University LinkedIn
• Nishant Sinha, OffNote Labs

Work done as part of Internship with OffNote Labs.

References

1. https://github.com/whai362/PSENet
2. https://github.com/Holmeyoung/crnn-pytorch

If this repository helps you, please star it. Thank you!