Thai OCR with deep learning

• For neural network crazes
• End-to-end training - only need images of characters and a list of characters in unicode - no need to provide stop marks or character cuts
• with Accuracy

1. Python 2.7.x
2. Cuda Toolkit 8.0 (https://developer.nvidia.com/cuda-toolkit)
3. TensorFlow r1.0 (https://www.tensorflow.org/install/)

1. Run python src/train.py to train.
2. Run python src/test.py to test.

1. Put new fonts under the "fonts" directory.
   • If fonts have different spacing diversity, the model will be hard to converge. (Need to go deeper?)

https://pvirie.wordpress.com/2016/11/23/a-challenge-to-test-neural-generative-attention/

MIT

That means free software.