Steven Rafaello Heydo Laksono
University of Technology Sydney

Published : 1 Documents Claim Missing Document
Claim Missing Document
Check
Articles

Found 1 Documents
Search

Comparative Analysis of Fully Connected Neural Network Architectures for Handwritten Hiragana Character Classification Steven Rafaello Heydo Laksono; Christian Arthur
International Journal of Information Engineering and Science Vol. 3 No. 2 (2026): May : International Journal of Information Engineering and Science
Publisher : Asosiasi Riset Teknik Elektro dan Infomatika Indonesia

Show Abstract | Download Original | Original Source | Check in Google Scholar | DOI: 10.62951/ijies.v3i2.265

Abstract

Handwritten character recognition remains an important challenge in computer vision, particularly for non-Latin writing systems that exhibit high variability in writing styles and stroke patterns. This study develops and evaluates fully connected neural network models for classifying handwritten Hiragana characters using the Japanese MNIST (KMNIST) dataset. The dataset consists of 70,000 grayscale images distributed across 10 character classes. Due to architectural constraints, each 28 × 28 image was transformed into a 784-dimensional feature vector and processed using fully connected neural networks with dropout regularisation. The primary objective of this study was to investigate how different architectural configurations influence classification performance while maintaining a minimum testing accuracy of 80% and reducing overfitting. Three experimental models were implemented, including a simple model, a wide model, and a deep model, representing variations in network width and depth. Model performance was evaluated using accuracy, precision, recall, F1-score, and training-validation loss analysis. The experimental results show that increasing model capacity improves classification performance; however, the type of architectural modification significantly affects generalisation. The simple model achieved 83.45% testing accuracy but exhibited substantial overfitting. The deep model improved performance to 90.99%, while the wide model achieved the best performance with 92.04% testing accuracy and the smallest generalisation gap. These findings demonstrate that wider architectures combined with appropriate dropout regularisation provide a better balance between accuracy and robustness than deeper architectures for this classification task.