Hiragana recognition

Handwritten hiragana recognition software.

Obtaining the ETL Character Database

To download ETL Character Database, you need to accept Conditions of Use and fill the Registration Form at etlcdb.db.aist.go.jp.

Obtaining unpack.zip

To download unpack.zip, go to http://etlcdb.db.aist.go.jp/file-formats-and-sample-unpacking-code.

Extracting hiragana dataset

• Clone this repository
• Download ETL Character Database and unzip into ETL directory
• Download unpack.zip and unzip into repository root folder
• Run create_dataset.sh script consisting of:
  • Virtual python environment creation and setup
  • ETL Character Database extraction
  • Hiragana dataset extraction
  • Minor samples cleanup
  • Samples size unification
  • Small characters removal

Classifier

Run train_hiragana_conv.sh to train your own CNN classifier, or use model.h5.
This model has 99.15% accuracy. Learning curves are shown below.

Character recognition

Run recognize.sh to launch the program.
To draw on canvas, hold left mouse button. To clear canvas, use CLEAR button.

Images