Lightweight Deep Learning Models for Facial Emotion Recognition: A Comparative Study

Facial emotion recognition has played a significant role in applications of artificial intelligence, enabling the detection and recognition of human facial emotions for diverse societal domains, including education, healthcare, security, and human-computer interaction. With the increasing demand for lightweight and efficient models suitable for real-world applications, the primary aim of this study is to develop a robust deep learning-based model that balances accuracy and computational efficiency, while recognizing facial emotions. To achieve this, EfficientNetB3, MobileNetV2, and their hybrid architecture were explored and optimized. EfficientNetB3 has been opted due to its resilient architecture and high accuracy, and selection of MobileNetV2 has been based on its lightweight configuration and suitability for edge and mobile devices. The hybrid architecture combines the strengths of both models to achieve an optimal trade-off between accuracy and efficiency. The models were evaluated using the Facial Emotion Recognition-2013 dataset, employing performance metrics i.e. accuracy, precision, F1-score, Root Mean Squared Error (RMSE), and Area Under the Curve (AUC). Experiments were conducted over 10 epochs with 3-fold cross- validation to enhance the generalization of the models. The results demonstrate that EfficientNetB3 and MobileNetV2 achieved accuracies of 95.29% and 94.01%, respectively, while the attained accuracy using hybrid architecture is 99.51%. These findings highlight the potential of advance facial emotion recognition using proposed models in real-world scenarios, particularly for lightweight and resource-constrained environments.