Deep learning framework for remote photoplethysmography signal quality classification using spectral gramian angular field

Remote photoplethysmography (rPPG) enables noncontact heart rate (HR) monitoring using facial videos; however, its accuracy is often compromised by motion artifacts and illumination variability. This study presents a robust signal quality assessment (SQA) framework that leverages spectral Gramian angular field (sGAF) representations and deep learning classifiers. Instead of using time-domain waveforms, the proposed method transforms the power spectral density (PSD) of rPPG signals into two-dimensional GAF images, which are then classified using Vision Transformer (ViT), ResNet18, and InceptionV3. A dataset was compiled using a camera under semi-practical indoor conditions, with synchronized contact PPG serving as the reference for ground-truth labeling. ViT combined with sGAF consistently demonstrated the best performance, exhibiting a classification accuracy of 96.2% and reducing HR estimation error (a mean absolute error of 1.78 bpm for 10-s segments). These findings validate the effectiveness of frequency-domain representations and transformer-based models in video-based physiological monitoring, particularly in scenarios where clean morphological waveforms are difficult to obtain.