kanji-classifier

OCR application that classifies almost 3000 Japanese kanji. Full list of characters can be accessed by running print_all_characters.py.
Deployed on kanji.al3xbro.me

Dependencies:

• Tensorflow, Keras: 2.12
• NumPy, OpenCV

Note: Follow the guide at https://www.tensorflow.org/guide/gpu to use your GPU for training

Performance:

• 92% accuracy for both validation and training sets.
• 0.18 training loss and 0.21 validation loss.

To Train:

1. Download an image dataset of your choice.
2. Modify the config.py file to contain the correct paths.
3. Run the image_preprocessing.py script to process images.
4. Run the delete_hiragana.py script to remove hiragana from the dataset.
5. Run the model_training.py script to train your model. Uses data augmentation to help the model generalize.

Testing:

• Run the predicting.py script to test your model.
• Try writing kanji in your own handwriting and testing your model on that. Have fun!

Resources:

• Datasets ETL8G and ETL9G from etlcdb were used for training and validation.
• Used etlcdb-image-extractor to extract images from these datasets. Thank you!