An Interpretable Deep Bayesian Model for Facial Micro-Expression Recognition

The Bayesian network is a powerful model for uncertain causal inference, but is limited to handle numerical data. In order to apply its excellent bidirectional inference ability to the image domain, this paper proposes an interpretable deep Bayesian model, which is based on deep learning technology to conduct semantic segmentation of facial micro-expressions and then extract features to construct the feature Bayesian network to analyze and infer causal relationships. Experiments show that the proposed model enables Bayesian networks to analyze image information, and enhances the interpretability of micro-expression recognition compared with deep learning models.