A Lightweight CNN for Multi-Class Classification of Handwritten Digits and Mathematical Symbols

Recognizing handwritten digits and mathematical symbols remains a nontrivial challenge due to handwriting variability and visual similarity among classes. While deep learning, particularly Convolutional Neural Networks (CNNs), has significantly advanced handwriting recognition, many existing solutions rely on deep, resource-intensive architectures. This study aims to develop a lightweight and efficient CNN model for multi-class classification of handwritten digits and mathematical symbols, with an emphasis on deployability in resource-constrained environments such as educational platforms and embedded systems. The proposed model, implemented in Julia using the Flux.jl library, features a compact architecture with only two convolutional layers and approximately 55,000 trainable parameters significantly smaller than typical deep CNNs. Trained and evaluated on a publicly available dataset of over 10,000 grayscale 28×28-pixel images across 19 symbol classes, the model achieves a test accuracy of 91.8% while maintaining low computational demands. This work contributes to the development of practical handwritten mathematical expression recognition systems and demonstrates the feasibility of using Julia for developing lightweight deep learning applications. Keywords - Digits, Mathematical Symbol, Classification, CNN