OCR-Kanji

A hobby project. A light CNN which detects and classifies japanese kanjis from different difficulty levels (N5, N4 etc). It contains a training data generator which creates japanese newspaper-like images with position and class labels.

Training data generation

First have a look through the parameters in the below file and run it in order to generate (recommended) about 1000 training images and labels.

generate_data.py

You can look at the generated labels for an example image by running:

visualise.py

Training the model

Model architecture is found in:

model1.py

Run:

train.py

It utilises a data generator defined in:

dataset.py

Predict

Run:

evaluate.py

The script takes the image defined in the file, splits it into patches to feed to the network and then it predicts the location and difficulty of the kanjis it found. It's not perfect.

Improvements

Probably a good idea to add katakana and hiragana classes. The training data generator should include larger characters, different fonts and diverse backgrounds.