Implementing Vision Transformer to Model Emotions Recognition from Facial Expressions

Emotions are essential to social interaction between interlocutors. It provides essential meaning to social interaction. It is important to recognise emotions along with verbal cues to fully understand the true meaning of a conversation in a social interaction setting. Several techniques can be applied to recognise emotions in the social interaction setting automatically. Emotions can be recognised and interpreted from the interlocutor's facial cues (e.g., facial expression recognition task) by using a sensor such as a camera. Recognising emotions from facial cues has several problems to be solved. This research aims to model robust emotion recognition from facial cues to recognise emotions (particularly from Asian people), as there is only limited research done in this area. This research aims to propose an emotion recognition model from facial cues by implementing Vision Transformers architecture. The dataset implemented in this research was explicitly performed by local (i.e., Indonesian) actors (The Indonesian Mixed Emotion Dataset - IMED). The results show that the proposed model can achieve the best testing accuracy of 100% with the best training loss of 0.0001, with 0.53 hours of training times for 50 epochs.