This tensorflow project shows how to use CNN to perform Chinese character recognition, a much more complicated task compared to digit recognition.

This code is based on a similar repo, but is cleaner and outperforms that one: This code defines a data iterator class to deal with the data preprocessing and use 2 different input pipeline for training and testing data.

Moreover, Training 16000 steps (~30 hour on my workstation with 8 CPU cores and 32 GB RAM) achieves a top 1 accuracy of 92.50% and a top 3 accuracy of 97.48%. If you have a GPU card, it should be much more faster.

The repo owner above also kindly shared the preprocessed dataset

python chinese_character_recognition_bn.py --mode=train --max_steps=16002 --eval_steps=100 --save_steps=500

python chinese_character_recognition_bn.py --mode=validation

Note that the network hasn't fully convergenced yet after 16000 mini-batches of training (though almost), so training for a longer time should be able to improve the performance furthur. As reported in this project report, the network has the potential to achieve a top 1 accuracy of 95% when properly trained (maybe even better, since in the aforementioned paper they didn't use batch normalization, which should improve generalization capacity of the network). So if you have a powerful GPU, run more training steps.

This is my checkpoint at step 16000. If you want to try it on your laptop or something else, feel free to do so!

I also attached the learning curve of chinese_character_recognition_bn.py: