Text recognition

Pytorch implementation for image-based sequence recognition tasks, such as scene text recognition and OCR.

Demo

demo images	VGG-BiLSTM-CTC	VGG-BiLSTM-CTC(case-sensitive)
	available	Available
	professional	professional
	londen	Fonden
	greenstead	Greenstead
	weather	WEATHER
	future	Future
	research	Research
	grred	Gredl

Performance

models	dataset	samples	accuracy	editing distance
VGG-BiLSTM-CTC	ICDAR 2013	1015	0.7438	0.4512
	ICDAR 2015	2077	0.4935	1.2335
	IIIT5k	3000	0.746	0.5493
	SVT	647	0.6831	0.5493
	Synth90k	891924	0.72	0.6146
VGG-BiLSTM-CTC	ICDAR 2013	1015	0.327	2.905
(case-sensitive)	ICDAR 2015	2077	0.4265	1.6364
	IIIT5k	3000	0.378	2.495
	SVT	647	0.4	2.3323
	Synth90k	891924	0.696	0.7671
ResNet-BiLSTM-CTC	ICDAR 2013	1015	0.3448	2.8916
(case-sensitive)	ICDAR 2015	2077	0.4535	1.546
	IIIT5k	3000	0.39	2.439
	SVT	647	0.4281	2.27
	Synth90k	891924	0.7188	0.6859

Usage

• Create lmdb dataset
  Explore create_lmdb_dataset.py for details
• Training and Evalution
  python train.py --adadelta --trainRoot {train_path} --valRoot {val_path} --batchSize 64 --cuda
python evaluate.py
• Demo
  Excute python web_demo.py to demo with web or python raw_demo.py to demo with single image

Training dataset

• Synth90k:
  • Introduction: The Synth90k dataset contains 9 million synthetic text instance images from a set of 90k common English words. Words are rendered onto natural images with random transformations and effects, such as random fonts, colors, blur, and noises. Synth90k dataset can emulate the distribution of scene text images and can be used instead of real-world data to train data-hungry deep learning algorithms. Besides, every image is annotated with a ground-truth word.
  • Link: Synth90k-download

Acknowledgements

This implementation has been based on this repository

Dependence

• lmdb
• FLask
• torch-summary
• CRAFT

Paper

An End-to-End Trainable Neural Network for Image-based Sequence Recognition and Its Application to Scene Text Recognition
Deep Residual Learning for Image Recognition
Character Region Awareness for Text Detection