Handwritten OCR for Kanji Characters

Open-source research project developing a CNN OCR (optical character recognition) dataset and model that can identify handwritten Kanji and other Japanese characters.

Dataset

Dataset is located at Machine_Learning/data. Each file is a PNG of the character in its filename. The original SVG dataset that was processed contains 7,000 images of handwritten kanji characters.

OCR Model

Model architecture is located at Machine_Learning/architecture.py. To import the model and weights in PyTorch run the following inside this cloned repo.

import torch
from Machine_Learning.architecture import KanjiNet

model = KanjiNet()
model.load_state_dict(torch.load('Machine_Learning/weights.pth')

Next Steps and Contributing

If you have improvements for data processing, training, or architecture please feel free to submit a pull request, any changes are welcome!

Next steps:

• [x] Apply random perspective transformation to training images for upsampling and unconstrained recognition of distorted characters
• [ ] Add segmentation and RNN to model for multi-character prediction

License

The Kanji-OCR project is open-source and is licensed under the MIT License.