Abstract
Thermal imaging offers a non-invasive and robust approach to emotion recognition by capturing facial temperature patterns that correlate with psychophysiological states. This study investigates the application of deep neural networks to classify six basic human emotions – happiness, sadness, fear, disgust, anger, and surprise – using facial thermograms. A balanced dataset was collected under controlled experimental conditions, and four deep learning architectures were evaluated: Convolutional Neural Network (CNN), Fully Convolutional Network (FCN), EfficientNet, and MobileNet. The models were trained and tested on a curated set of preprocessed thermal facial images. Among the evaluated architectures, FCN achieved the highest classification accuracy of 90.04%. The results demonstrate that deep learning models, particularly FCNs, are well-suited for emotion recognition from thermal data, with potential applications in psychophysiological monitoring, healthcare, and real-time humancomputer interaction systems.