Evaluating the Performance of Deep Learning Models in Handwritten Digit Recognition
Madan Lal Saini, Bharat Tripathi, Mohammed Sohail Mirza
Abstract
Handwritten digit recognition is an essential requirement in the field of computer vision, with applications ranging from processing bank checks to automated the postal system to digitization of historical documents. Different deep learning techniques are now being used to identify handwritten numbers. In order to recognize handwritten digits, this study compares and contrasts four well-known deep learning architectures: LeNet-5, AlexNet, VGG-16, and ResNet-10l. In the experimental study, handwritten digit datasets from MNIST, DIDA, and MNIST MIX were used. The aim of this study is to identify the optimal algorithm that can provide a respectable level of accuracy. This research gives a succinct review of each architecture, focusing its distinctive design traits and advantages. The dataset is preprocessed to ensure uniformity and to enhance model generalization. The evaluation is conducted using standard metrics such as accuracy, precision, recall, and Fl-score, as well as a visual examination of the model's predictions. According to the results, all four architectures achieved remarkable accuracy than traditional machine learning techniques. However, the computational complexity, training duration, and generalization abilities are having notable differences. LeNet and ResNet are showing good accuracy and AlexNet and VGG are having quicker training times.