Facial Expression Recognition (FER) is a pivotal technology in human-computer interaction, with applications spanning psychology, virtual reality, and advanced driver assistance systems. Traditional FER using visible light cameras faces challenges in low light conditions, shadows, and reflections. This study explores thermal imaging as an alternative, leveraging its ability to capture heat radiation and overcome lighting issues. We propose a novel approach that combines pre-trained models, particularly EfficientNet variants, with Grey Wolf Optimization (GWO) and various classifiers for robust emotion recognition. Ten pre-trained CNN models, including variants of EfficientNet (EfficientNet-B0, B3, B4, B7, V2L, V2M, V2S), ResNet50, MobileNet, and InceptionResNetV2, are utilized to extract features from thermal images. GWO is employed to optimize the parameters of four classifiers: Support Vector Machine (SVM), Random Forest, Gradient Boosting, and k-Nearest Neighbors (kNN). The proposed approach is evaluated on two well-known thermal image datasets: IRDatabase and KTFE. The highest accuracy achieved on the IRDatabase dataset is 91.42%, using the combination of EfficientNet-B7 with GWO and either kNN or SVM for eight different emotions (Fear, Anger, Contempt, Disgust, Happy, Neutral, Sad, and Surprise). On the KTFE dataset, the highest accuracy is 99.48%, achieved by combining EfficientNet-B7 with GWO and Gradient Boosting for seven different emotions (Anger, Disgust, Fear, Happy, Neutral, Sadness, and Surprise). These results demonstrate the robustness and efficiency of the proposed method in thermal image-based emotion recognition, making it a promising solution for overcoming the limitations of traditional visible light-based FER systems.

Keywords: Emotion Recognition, Thermal Images, Deep Learning, Transfer Learning, Gray Wolf Optimization, EfficientNet-B7, Support Vector Machine, k-Nearest Neighbors, Gradient Boosting.