Implementation of Convolutional Neural Network for Hiragana Classification

Abstract

Hiragana is one of three writing systems in Japanese that is widely used in various texts, from literature to everyday media. Hiragana characters consist of 46 basic symbols, each representing one syllable. Hiragana character classification is a complex challenge in image processing and machine learning. Despite various advances, achieving high accuracy in Hiragana character classification remains difficult. This research aims to increase the accuracy of Hiragana character classification by applying a Convolutional Neural Network (CNN). The CNN method was chosen because of its strong feature extraction and image classification capabilities. This research focuses on the optimizers, including Adam, SGD, RMSprop, Adamax, Adagrad, and Adadelta, for Hiragana character classification. The dataset used in this research focuses on ten Hiragana characters: Aa, Ki, Su, Tsu, Na, Ha, Ma, Ya, Re, and Wo. The research results show that using the Adam optimizer in the CNN model training process achieved the highest accuracy of 97.37%. Model performance analysis also considers important metrics such as precision, recall, and F1-Score to validate the effectiveness of the Adam optimizer in improving model performance. These results emphasize the role of the Adam optimizer in improving the accuracy of image classification models. This research significantly contributes to Hiragana character classification by highlighting the importance of choosing the right optimizer to improve model performance. In conclusion, the Adam optimizer selection proved effective in increasing the accuracy of Hiragana character classification using CNN.