Microexpression Recognition from Action Unit-Variational Graph Autoencoder

Emotion recognition has helped humans express their feelings, assisting in therapy sessions or criminal investigations. However, humans are clever enough to hide their feelings, making them hard to recognize. Micro-expression can help reveal their genuine emotions from very fast-changing expressions. This study attempts to recognize micro-expression based on a deep learning model. A recent related study used action unit with graph neural networks but mostly fail to generalize the facial graph structure. To overcome the issue, we propose an integrated deep learning model from action unit-graph neural networks to build the facial graph structure and variational autoencoders to extract the latent information from it called AU-VGAE. In such way, we offer better generalization to learn the facial graph structure. In this study, the proposed model was evaluated on a public micro-expression dataset. The evaluation showed the proposed model performs better than state-of-the-art models.