This is a demo using lstm and ctc to recognize a picture of a series numbers with blanks all at once.
For example:given the piture below the model would give result 73791096754314441539
.
pip install opencv-python
brew install cmake
brew tap homebrew/science
brew install opencv
sh ./prepare_train_data.sh
The prepare_train_data.sh
script would download the SUN database and extract the pitures to bgs dir. Then you can run python gen.py
to generate test and train dir.
When the train and test data set are ready you can start the train process by nohup python lstm_and_ctc_ocr_train.py
.
- Python 2.7+
- Tensorflow 1.0+
- python_speech_features
- numpy
- scipy
This project is licensed under the terms of the MIT license.
See README for more information.