Divided Block Multiscale Convolutional Network for Micro-expression Recognition

Micro-expression (ME) is a subtle change in the face, which can be used to judge human subjective feelings. It has broad application prospects in medical diagnosis and business negotiation. However, due to the complexity of ME muscle movements and the lack of ME trainable data, the research of micro-expression recognition (MER) still faces a series of challenges. In this paper, we propose a new divided block multiscale convolution network (DBMNet), which could learn from four different optical flow (OF) feature images obtained between the onset and apex frames of ME samples. Through the proposed block-divided multiscale convolution module (BMCM), more detailed and useful multiscale advanced features of ME could be effectively extracted. In order to better address the problem of class imbalance on the ME dataset, this paper uses the weighted cross entropy (CE) loss function, which could obviously alleviate the impact of class imbalance. Finally, 5-class experiment is conducted on the composite dataset to show that the proposed method has superior performance and is comparable to those of the most advanced methods.