A implement of paper ：https://arxiv.org/abs/1904.01941
This is the result of training 200,000 steps.
Result
Useage
Down ckpt https://drive.google.com/file/d/1-AL-T3-WucUZsWMUHqyWovuXEvqYZCi9/view?usp=drivesdk
Train: Check your syntext dataset dir in datagen.py, and change train(False) to train(True) at last line. Run craft.py
text:
modify below:
if __name__ == "__main__":
train(False)
# test('ckpt_path', 'text_image_path')
TO
if __name__ == "__main__":
# train(False)
test('ckpt_path', 'text_image_path')
The first argument is your .ckpt path, seconed argument is text image path.
Fine-tuning:
TO DO
If you want to do this, you could use train(False), it means you load a checkpoint then training
Weak supervision:
TO DO