Micro-Expression Recognition Based on Multiple Aggregation Networks

Micro-expression is a low-intensity, short-term spontaneous facial activity that can reflect people's true feelings. Existing methods mainly extract hand-crafted descriptors from the whole face, which are not enough to capture detailed information of the local regions and are not optimal due to depending on the experience of researcher. Thus, we propose a multiple aggregation networks to explore the impact of local facial regions on micro-expressions recognition in detail. The framework uses multiple different network branches to extract frame-level information about the facial regions of interest, as well as the holistic features of whole face. Finally, the physical meaning of statistical parameters is fully utilized to characterize the average and dynamic changes of frame-level features to generate video-level features. Finally, use video-level features as the input of SVM for micro-expression classification. Experiments are conducted on CASME, CASME II and SMIC databases. The results demonstrate that the proposed method is superior to previous works.