A Deep Learning Approach to Generate Thermal Images Synthetically for Emotion Recognition Applications

Abstract

Thermal imaging is widely used in emotion recognition for its ability to capture physiological responses such as temperature variations in facial regions. However, collecting large-scale thermal datasets is challenging due to cost, privacy concerns, and limited availability. This presents a deep learning approach to synthetically generate thermal images from RGB images for emotion recognition applications. A Generative Adversarial Network (GAN)-based model is trained to make it learn the mapping between visible and thermal domains, ensuring realistic thermal image synthesis. The proposed method enhances existing emotion recognition systems by providing augmented thermal datasets, reducing dependency on expensive thermal cameras. Experimental results shows that using synthetic thermal images enhances the accuracy of deep learning-based emotion recognition models.