A New Model to Detect COVID-19 Coughing and Breathing Sound Symptoms Classification from CQT and Mel Spectrogram Image Representation using Deep Learning

Abstract

Deep Learning is a relatively new Artificial Intelligence technique that has shown to be extremely effective in a variety of fields. Image categorization and also the identification of artefacts in images are being employed in visual recognition. The goal of this study is to recognize COVID-19 artefacts like cough and also breath noises in signals from real-world situations. The suggested strategy considers two major steps. The first step is a signal-to-image translation that is aided by the Constant-Q Transform (CQT) and a Mel-scale spectrogram method. Next, nine deep transfer models (GoogleNet, ResNet18/34/50/100/101, SqueezeNet, MobileNetv2, and NasNetmobile) are used to extract and also categorise features. The digital audio signal will be represented by the recorded voice. The CQT will transform a time-domain audio input to a frequency-domain signal. To produce a spectrogram, the frequency will really be converted to a log scale as well as the colour dimension will be converted to decibels. To construct a Mel spectrogram, the spectrogram will indeed be translated onto a Mel scale. The dataset contains information from over 1,600 people from all over the world (1185 men as well as 415 women). The suggested DL model takes as input the CQT as well as Mel-scale spectrograms derived from the breathing and coughing tones of patients diagnosed using the coswara-combined dataset. With the better classification performance employing cough sound CQT and a Mel-spectrogram image, the current proposal outperformed the other nine CNN networks. For patients diagnosed, the accuracy, sensitivity, as well as specificity were 98.9%, 97.3%, and 98.1%, respectively. The Resnet18 is the most reliable network for symptomatic patients using cough and breath sounds. When applied to the Coswara dataset, we discovered that the suggested model's accuracy (98.7%) outperforms the state-of-the-art models (85.6%, 72.9%, 87.1%, and 91.4%) according to the SGDM optimizer. Finally, the research is compared to a comparable investigation. The suggested model is more stable and reliable than any present model. Cough and breathing research precision are good enough just to test extrapolation as well as generalization abilities. As a result, sufferers at their headquarters may utilise this novel method as a main screening tool to try and identify COVID-19 by prioritising patients' RT-PCR testing and decreasing the chance of disease transmission.