Handwritten Digit Recognition Using Deep Learning

Abstract

Handwritten digit recognition (HDR) remains one of the most fundamental and widely explored tasks in pattern recognition and computer vision, serving as a benchmark for evaluating machine learning and deep learning architectures. This study presents a comparative evaluation of three neural network models—Simple Perceptron (SP), Multilayer Perceptron (MLP), and Convolutional Neural Network (CNN) for handwritten digit classification using the standard MNIST dataset. The methodology includes data normalization, reshaping, and one-hot encoding to ensure consistent preprocessing across models. The SP model achieved a baseline accuracy of 92.7%, demonstrating limited capacity to capture complex spatial relationships. The MLP model, incorporating two hidden layers with nonlinear activations, improved recognition accuracy to 97.2%, confirming the benefit of deeper, fully connected representations. The CNN model achieved the highest validation accuracy of 98.9% by leveraging convolutional and pooling layers to efficiently extract local spatial hierarchies. Training and validation performance analyses confirmed stable convergence and minimal overfitting. Overall, the results demonstrate that CNNs significantly outperform classical and shallow neural architectures for HDR tasks, offering superior generalization, computational efficiency, and suitability for real-world applications in optical character recognition (OCR) and document automation.