GAT-Net: A Network using Grid Attention and Transformer for Dynamic Facial Expression Recognition

Dynamic facial expression recognition (DFER) in the wild has received widespread attention recently. There are complex factors such as face occlusion and pose variation in the wild. Facial expression recognition has a subtle competition between capturing local features of a human face and obtaining a global feature representation. This paper proposes an end-to-end DFER network GAT-Net based on the grid attention module and Transformer, which improves the robustness and accuracy of DFER in the wild. Specifically, GAT-Net is divided into two components: spatial feature extraction and temporal feature processing. The grid attention module of the spatial feature extraction component guides the network to pay attention to the local salient features of the face, which reduces the interference of field occlusion and non-frontal poses. The Transformer in the temporal feature processing component guides the network to learn the temporal relationship of high-level semantic features and the global representation of facial expression features. These two components balance the subtle competition between local features and global feature representations of facial expressions. The ablation experiment has proved the effectiveness of the grid attention module and Transformer. Experiments demonstrate that our GAT-Net outperforms state-of-the-art methods on DFEW and AFEW benchmarks with accuracies of 67.53%, and 50.14% respectively.