foamliu / CRNN

PyTorch re-implementation of CRNN: Convolutional Recurrent Neural Network

Convolutional Recurrent Neural Network

Introduction

This is a PyTorch re-implementation of CRNN: Convolutional Recurrent Neural Network (paper). The features are summarized blow:

DataSet

We use the synthetic dataset (mjsynth) released by Jaderberg et al. as the training data. The dataset contains 8 millions training images and their corresponding ground truth words. Such images are generated by a synthetic text engine and are highly realistic.

@article{Jaderberg14c,
      title={Synthetic Data and Artificial Neural Networks for Natural Scene Text Recognition},
      author={Jaderberg, M. and Simonyan, K. and Vedaldi, A. and Zisserman, A.},
      journal={arXiv preprint arXiv:1406.2227},
      year={2014}
    }

Dependency

• PyTorch 1.1.0

Usage

Data Pre-processing

Extract training & test images:

$ python extract.py

Train

$ python train.py

If you want to visualize during training, run in your terminal:

$ tensorboard --logdir runs

Demo

Pick 10 random examples from test set of mjsynth:

$ python demo.py

Image	Word
	U---n-r-e--a-l--is-t-i---c => Unrealistic
	O---V-E-R--R-U--L--E-----S => OVERRULES
	W---E-S--T-M-IIN-S-T-E---R => WESTMINSTER
	m-------o----r--ee-l-----s => morels
	P--E-R-P-L-E-A-I-T-A--R--S => PERPLEAITARS
	E----------D-------------P => EDP
	G-----e-o-lo--giis-t-----s => Geologists
	r-----e----b---a----t---ee => rebate
	F--O-R-E-D-O--O-M--IIN---G => FOREDOOMING
	u--n--re-p-r-es-e-n-te---d => unrepresented