This project refers to the handwritten recognition with CNN and RNN, decode with CTC.
- [Model1 - word_model.png] Train on WORD unit of dataset.
- [Model2 - line_model.png] Train on LINE unit of dataset.
Test on IAM dataset:
Model | Test Unit | Number of samples | CER(%) | WER(%) |
---|---|---|---|---|
WORD | WORD | 19289 | 10.39 | 26.97 |
WORD | LINE | 2192 | 21.73 | 46.00 |
LINE | LINE | 2192 | 08.32 | 28.99 |
[Google colab]
$ python3 Train.py
$ python3 Prediction.py
$ python3 EvaluationWord.py
$ python3 EvaluationLine.py
https://towardsdatascience.com/build-a-handwritten-text-recognition-system-using-tensorflow-2326a3487cd5 http://norvig.com/spell-correct.html