Micro-Expression Recognition Using Micro-Variation Boosted Heat Areas.

Micro-Expression Recognition has been a challenging task as transitory micro-expressions only appear in a few of facial areas. In this work, we aim to improve the recognition performance by boosting the micro-variations in the learned areas. This paper proposes an architecture, deriving facial micro-variation heat areas and then integrating in conjunction with the micro-expression recognition network, to learn the micro-expression features in an end-to-end manner. The method is constructed by the Heat Areas Estimator from Heatmap (HAEH), which is to produce micro-variation heat areas as facial geometric structure, and the Temporal Facial Micro-Variation Network (TFMVN) for learning the fusion features. The method can define and capture facial heat areas significantly contributed to the micro-expressions. Our approach activates or deactivates corresponding feature maps from the heat areas to guide feature learning. We perform experiments on CASME II dataset and SAMM dataset. The experimental results show that we achieve state-of-the-art accuracy, and the method demonstrates good generalization ability for cross-dataset. Moreover, we validate three pivotal components’ effectiveness within our architecture.