laugha / OCRetina

End-to-end one-stage CNNs for text detection and recognition based on RetinaNet

OCRetina

THIS REPO IS UNDER DEVELOPMENT

Solutions

This repo provide end-to-end CNNs for text detection and recognition based on RetinaNet. Currently, two solutions are proposed:

Two-stage OCR system

Facebook's Rosetta system for text detection and recognition in images (source: Facebook [3])

One-stage OCR system

One-stage architecture for text detection and recognition in images. Image is adapted from Rosetta's paper (Sorry I'm not good at drawing)

References

[1] RetinaNet: https://arxiv.org/abs/1708.02002

[2] CTC loss: http://citeseerx.ist.psu.edu/viewdoc/summary?doi=10.1.1.139.5852

[3] Rosetta: https://research.fb.com/wp-content/uploads/2018/10/Rosetta-Large-scale-system-for-text-detection-and-recognition-in-images.pdf

Based sources code

This repo was built using several materials as below.

Keras RetinaNet: https://github.com/fizyr/keras-retinanet

Keras OCR: https://github.com/keras-team/keras/blob/master/examples/image_ocr.py