A cross-database micro-expression recognition framework based on meta-learning

Micro-expressions are facial expressions that are revealed unconsciously when suppressing true emotions and are widely used in multiple tasks, such as deception detection. However, at present, the amount of available micro-expression data is small, and there are large differences between different databases, so it is still difficult to accurately perform cross-database micro-expression recognition, which hinders real applications of recognition. To address this issue, this article first presents a meta-learning framework suitable for cross-database micro-expression recognition named Meta-CDMERF, which is trained by combining multiple micro-expression databases. Then, a residual feature-wise linear (RFL) module is proposed to generate more feature distributions and adaptively choose representative features during multi micro-expression database training, thereby reducing the feature distance between samples of the same type in different databases. Next, a new loss function is designed, which combines the cross-entropy loss function with an interclass loss function. Specifically, the inter-class loss is based on the mean value of similar features from the support set and aims to increase the distance between samples of different categories, thereby capturing subtle changes in micro-expression images. Finally, the unweighted average recall (UAR) and unweighted F1 score (UF1) values of the proposed method on the CASME II database are 62.64% and 60.00%, respectively, reaching state-of-the-art performance.