What is this?

Japanese hand written Hiragana (ひらがな) recoginition by deep learning. Accuracy is 99.78%.

Papers

My implementation is based on Charlie Tsai's "Recognizing Handwritten Japanese Characters Using Deep Convolutional Neural Networks". http://cs231n.stanford.edu/reports2016/262_Report.pdf

This is just a VGG-like convnet. https://arxiv.org/abs/1409.1556

What I modified

• Resized image to 32x32 px. Charlie Tsai's paper is 64x64 px.
• Initial weights are from a normal distribution with 0.1 standard deviation.
• Used AdaDelta and increased the number of epochs to 400.
• Add random rotation and zoom to images.

By these changes the accuracy increased from 96.13% to 99.78%.

Librarys

You need following librarys.

• Anaconda (Python 3.5)
• Theano 0.8.2 or TensorFlow 0.10.0
• Keras 1.1.0
• scikit-learn 0.18.0 (included in Anaconda)
• CUDA 7.5
• cuDNN 5.0

Config files

~/.theanorc

[global]
floatX = float32
device = gpu

[lib]
cnmem = 1

[nvcc]
flags=-D_FORCE_INLINES

~/.keras/keras.json

Theano backend

{
    "epsilon": 1e-07,
    "image_dim_ordering": "th",
    "floatx": "float32",
    "backend": "theano"
}

or TensorFlow backend.

{
    "epsilon": 1e-07,
    "image_dim_ordering": "tf",
    "floatx": "float32",
    "backend": "tensorflow"
}

Modify both image_dim_ordering and backend. Theano 0.8.2 is faster than TensorFlow 0.10.0.

Dataset

Please download dataset from http://etlcdb.db.aist.go.jp/?page_id=651 and extract to ETL8G folder. Dataset contains 160 person hiragana characters.

How to run

Just run python learn.py.