Hierarchical Approaches to Handwritten Digit Recognition: A Study of Modern Neural Networks

Handwritten digit recognition has several applications in multiple industries in this modern era for enhancing efficiency, accuracy, and accessibility. Convolution Neural Networking (CNNs) have emerged for the precise result of this task since it is a powerful tool for this task due to the ability to learn hierarchical features from data. Several architectures of convolution neural networking (CNNs) are used in this field. This study investigated the efficiency of CNN architectures in recognizing handwritten digits. The considered architectures are VGG16, ResNet50, and DenseNet. However, DenseNet stands out with the highest efficiency of 99.19% compared to traditional architectures like VGG16 and ResNet50. A benchmarked dataset of the machine learning field, MNIST, has been used for training data. Through experimental evaluation, it was observed that the three CNN architectures mentioned achieved high accuracy rates on the MNIST dataset, namely 95.9% for VGG16, 98.5% for ResNet50, and 99.19% for DenseNet. However, DenseNet has been proven to be the most accurate.