Development of an Application for Studying Kanji Using a Neural Network

The topic of the present study is aimed at investigating some key questions concerning convolutional neural network (CNN, or ConvNet) implementation for recognizing handwritten Japanese hieroglyphs. Therefore, this is a complex classification task in deep learning. The objective of the project is to write from scratch some realization of a convolutional neural network in Python, using deep learning library Keras running on top of TensorFlow backend. The trained neural network will be a part of the web application where the user can learn Japanese characters using the canvas of JavaScript to draw the hieroglyphs for recognition. In order to improve deep learning performance CNN has been experimented to increase the accuracy of a network model; also, accuracy has been analyzed on the validation data set during the learning process in the following research. The network has been trained using the ETL Character Database. In all classification tasks, convolutional neural networks could be able to achieve high recognition rates.