Compare the classification performances of convolutional neural networks and capsule networks on the Coswara dataset

Abstract

Since the beginning of the COVID-19 pandemic, researchers have developed numerous machine learning models to distinguish between positive and negative COVID-19 sounds. The aim of this study is to compare the classification performances of convolutional neural networks (CNN) and capsule networks (CapsNet) on the Coswara dataset, which includes 1404 healthy subjects and 522 COVID-19 positive subjects, each containing nine different types of sounds. The dataset was preprocessed by using oversampling and normalization techniques after feature extraction. k-fold cross-validation was used (where k=10) to train and evaluate the models. The CNN classifiers achieved a 94% ACC, while the CapsNet classifiers achieved an 90% ACC. Furthermore, when using leave-one-out cross-validation, the CNN classifier achieved an ACC of 99%. we also compared the performance of the CNN and CapsNet networks on the Coswara dataset without preprocessing. Without oversampling techniques, the CNN classifiers achieved an 93% ACC, compared to 54% for the CapsNet classifiers. When normalization techniques were not applied, the CNN classifiers achieved an 86% ACC, while the CapsNet classifiers achieved a 26% ACC.