Comparative Analysis of Deep Learning Models for Handwritten Character Recognition Using the EMNIST Dataset

Abstract

This report presents a study where handwritten character recognition has been done using the EMNIST dataset that comprises 62 classes of letters and digits. The aim of this project is to classify the characters using three different deep learning models: Multi-Layer Perceptron (MLP), Convolutional Neural Network (CNN), and CNN with Residual Connections. They were realized, deployed, and assessed to compare their outcomes. The findings portray that the CNN with Residual Connections archive the highest precision. This ability to capture complex features better than the MLP and traditional CNN models is the motivation for this approach. The easiest path in this work is finding a simple solution to the issues arising in the transcripts, such as overfitting and tuning hyperparameters. At the same time, recommendations for the resolution of these problems and prospective improvements are provided.