Facial Expression Recognition via Deep Action Units Graph Network Based on Psychological Mechanism

Abstract

Facial expression recognition (FER) is currently a very attractive research field in cognitive psychology and artificial intelligence. In this paper, an innovative FER algorithm called deep action units graph network (DAUGN) is proposed based on psychological mechanism. First, a segmentation method is designed to divide the face into small key areas, which are then converted into corresponding AU-related facial expression regions. Second, the local appearance features of these critical regions are extracted for further action units (AUs) analysis. Then, an AUs facial graph is constructed to represent expressions by taking the AU-related regions as vertices and the distances between each two landmarks as edges. Finally, the adjacency matrices of facial graph are put into a graph-based convolutional neural network to combine the local-appearance and global-geometry information, which greatly improving the performance of FER. Experiments and comparisons on CK+, MMI, and SFEW data sets reveal that the DAUGN achieves more competitive results than several other popular approaches.