Comparative Analysis of Lightweight CNNs for Multilingual Handwritten Character Recognition

Abstract

Handwritten character recognition remains a fundamental problem in pattern recognition, with applications ranging from digital document processing to historical manuscript digitization. Multilingual handwritten character recognition introduces additional challenges due to variations in scripts, diacritics, and handwriting styles. In this work, we explore lightweight convolutional neural networks (CNNs) for recognizing handwritten characters across multiple languages using the merged T-H-E dataset, which combines visually identical upper- and lower-case characters into 54 classes. Four CNN architectures, as LeNet-5 (baseline), Simple CNN, Depthwise-CNN, and MiniVGGNet, are evaluated based on their efficiency, number of parameters, training time, and misclassification behavior. Experiments were conducted under standardized hyperparameters and using a fixed random seed to ensure a fair comparison. Our analysis shows that lightweight networks can achieve competitive performance while requiring minimal computational resources, and the study provides insights into common misclassification patterns related to visually similar characters and diacritics. These findings establish a fast and resource-efficient baseline for multilingual handwritten character recognition on small binary datasets.