Handwriting Recognition of Hiragana Characters using Convolutional Neural Network

Abstract

Hiragana is one of the basic types of letters used in Japanese writing. This research proposes the method of recognizing Hiragana's writing characters using the Convolutional Neural Network (CNN) method. At the preprocessing stage, the segmentation process is carried out using the thresholding method to segment, followed by the process of noise removal, resize, and cropping for image normalization. In the CNN training process, maxpooling methods and danse functions are used for the fully connected process. Whereas in the testing phase the accuracy of using the Adam Optimizer tool. By using 1000 image datasets consisting of 50 characters, each with 50 samples, and with a composition of 950 training data and 50 testing data, the accuracy is 95%. This proves that the CNN method has a good performance for Hiragana character recognition.