This code implements recognition part of the ASTER
The text recognizer is an sequence to sequence model based on attention. The origin paper can be found at Here
python3 train.py [train_res_path] [valid_res_path] [model_path]
python3 test.py [test_res_path] [model_path]
I trained the model based on Invoice dataset which contain 1830 vocabularies and 18710 training images and 6752 validation images. The final accuracy on the validation data is about 70.2% on line base accuracy and 90.1% on word base accuracy.