Cross-database micro-expression recognition with deep convolutional networks

Micro-expression recognition (MER) is attracting more and more interests as it has important applications for analyzing human behaviors. Since the recognition ability for individual datasets has been improved greatly, few works have been devoted to the cross database task of MER, which is more challenging for capturing the subtle changes of micro-expressions from different environments. In this paper, we employ an end-to-end deep model for learning the representation and classifier automatically. In the deep model, the recurrent convolutional layers are utilized to exploit the learning ability with the optical flow fields of micro-expression sequences, which are enhanced by a motion magnification procedure. To ease the influence of samples from different datasets (environments), we present three normalization methods (i.e., sample-wise, subject-wise and dataset-wise methods) to restrain the variations of samples. The experiments are performed on the cross database of MERC2019 challenge, and achieve comparative performance than the baseline method.