Adaptive and Compact Graph Convolutional Network for Micro-expression Recognition

Micro-expression recognition (MER) is a very challenging task since the motion of Micro-expressions (MEs) is subtle, transient and often occurs in tiny regions of face. To build discriminative representation from tiny regions, this paper proposes a novel two stream network, namely Adaptive and Compact Graph Convolutional Network (ACGCN). To be specific, we propose a novel Cheek Included Facial Graph to build more effective structural graph. Then, we propose the Tightly Connected Strategy to adaptively select structural graph to build compact and discriminative facial graph and adjacency matrix. We design the Small Region module to enlarge the interested feature in tiny regions and extract effective feature to build strong and effective node representations. We also adopt the spatial attention to make the network focus on the visual feature of salient regions. Experiments conducted on two micro-expressions datasets (CASME II, SAMM) show our approach outperforms the previous works.