Deep learning based MobileNet and multi-head attention model for facial expression recognition

Facial expressions is an intuitive reflection of a person’s emotional state, and it is one of the most important forms of interpersonal communication. Due to the complexity and variability of human facial expressions, traditional methods based on handcrafted feature extraction have shown insufficient performances. For this purpose, we proposed a new system of facial expression recognition based on MobileNet model with the addition of skip connections to prevent the degradation in performance in deeper architectures. Moreover, multi-head attention mechanism was applied to concentrate the processing on the most relevant parts of the image. The experiments were conducted on FER2013 database, which is imbalanced and includes ambiguities in some images containing synthetic faces. We applied a pre-processing step of face detection to eliminate wrong images, and we implemented both SMOTE and Near-Miss algorithms to get a balanced dataset and prevent the model to being biased. The experimental results showed the effectiveness of the proposed framework which achieved the recognition rate of 96.02% when applying multi-head attention mechanism