SEDL: Learning Emotion Dynamics from Facial Representations Using Self-Supervised Approaches

Abstract

Facial Expression Recognition (FER) is an essential component of affective computing, with significant applications in healthcare, behavioral analysis, and assistive technologies for neurodiverse and mentally challenged individuals. Despite considerable progress, traditional machine learning and supervised deep learning approaches are often constrained by their dependence on large labeled datasets and their limited ability to capture the temporal dynamics of emotional expressions. To address these challenges, this paper proposes a novel Self-Supervised Emotion Dynamics Learning (SEDL) framework that integrates contrastive self-supervised learning with temporal emotion progression modeling. The proposed approach enables the learning of meaningful feature representations from unlabeled facial images while simultaneously capturing the evolution of emotional states over time. This combination enhances the model’s ability to generalize across diverse and real-world conditions. The framework is evaluated on a dataset of facial expressions from neurodiverse individuals, demonstrating its applicability in practical and sensitive environments. Comparative analysis with traditional machine learning, supervised deep learning, and self-supervised approaches indicates that the proposed method provides improved performance and robustness. Overall, the proposed SEDL framework offers a scalable and efficient solution for emotion recognition, addressing key limitations of existing FER systems. It has strong potential for deployment in real-time applications such as behavioral monitoring, mental health assessment, and intelligent assistive systems.