Ancient Japanese Kana classification with CNN

Install venv

$ python -m venv <env_name>
$ source <env_name>/bin/activate
(<env_name>)$ pip install -r ./requirements.txt

on Windows, run <env_name>\Scripts\activate.bat instead of source <env_name>/bin/activate

Training model

Run

$ python main.py -d [data_path]

For example

$ python main.py -d /alcon2019/dataset/train_kana

As default, all images are read and preprocessed to generate ./preprocessed_data.npy If you modify the model structure and want to retrain again, while the data are the same, you can run

$ python main.py -d [data_path] -r False

to skip the preprocessing. The program will import processed data from preproccessed_data.npy directly.

After training, the model is stored in ./model/ and the label binarization object is stored in ./labelbin/. These two files are used for prediction of new input image.

A visualized result of the loss and accuracy rates over each epoch is stored in ./plot/.

A json file counting the incorrect validations of each class is stored as ./incorrectCount.json

Classify new input image with the trained model

From terminal

Run

$ python classify.py -m ./model/model -l ./labelbin/labelbin -i [image_path]

to classify a Kana image.

From code

from classify import KanaClassifier
import cv2

model_path = './model/model'
labelbin_path = './labelbin/labelbin'
image_path = 'some_path'

kc = KanaClassifier(model_path, labelbin_path)
img = cv2.imread(image_path, cv2.IMREAD_GRAYSCALE)
result = kc.classify(img)

result is a list of dictionaries containing two keys, unicode and confidence, sorted by the confidence from high to low.