SCL-FExR: supervised contrastive learning approach for facial expression Recognition

Facial Expression Recognition (FER) is a significant field of computer vision and has emerged as a crucial component of Human-computer interaction. Breakthroughs in self-supervised representation learning have resulted from a renaissance of work in contrastive learning, following the state-of-the-art performance in unsupervised training of deep image models. However, due to the random sampling of false negatives for contrastive loss calculation, the representation quality might degrade in FER. In this work, we extend the self-supervised contrastive learning technique to the fully supervised setting to effectively exploit label information in classifying facial expressions. Therefore, we propose a Supervised Contrastive Learning- Facial Expression Recognition (SCL-FExR) system to create a model which is robust for real-world emotion detection. Our goal is not to compete with the highly complex state-of-the-art CNN-based Deep Neural Network, but to establish a method that can be incorporated to achieve similar performance but with less-complex models and more robustness. We demonstrate the effectiveness of the suggested method using three FER datasets: FER2013, AffectNet, and CK+. On FER2013, we achieved a similar accuracy of 76%, establishing a method that can be incorporated into less complex CNN-based Deep Neural Networks to achieve robustness and be significantly more noise-resistant. The secondary aim is to show how a data-based strategy may be used to train very complicated deep learning models instead of a model-based approach, which solves the issue of computational expenditure.