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