Improving the quality of respiratory signals extracted from the segmented mask area

Abstract

The COVID-19 pandemic has underscored the importance of wearing facial masks and monitoring respiratory health to prevent the spread of the virus. In this study, we developed a model for segmenting facial masks in thermal images. We applied the model to segment face masks in different conditions, including a person walking toward the observing camera. The segmented regions were further processed using different erosion masks to analyze the influence of the selected sources on the quality of the estimated respiratory signals. The Signal-to-Noise Ratio (SNR) was used as a quality measure. Additionally, the extracted respiratory signals were compared with two reference signals: binary signals generated by participants who signaled the inhalation phase and pressure signals measured with a respiratory belt. Our findings show a high level of concordance between the respiratory signals derived from the segmented mask region and those from the respiratory belt, validating the effectiveness of thermal imaging for capturing respiratory patterns. Notably, the signal-to-noise ratio (SNR) was higher for the segmented mask than the detection methods used in previous works. Specifically, for the mask segmentation task, the mean SNR improved by 4.3 compared to facial mask detection. The segmentation model achieved a mean Average Precision (mAP) of 0.992 for segmentation tasks and 0.857 mAP at the 50–95 % threshold using the Yolov8 “nano” architecture. This study underscores the potential of thermal imaging for non-invasive respiratory monitoring and highlights the explainability and accuracy of selecting the facial mask region for signal extraction.