Dynamic Expression Recognition Based on Attention Mechanism and Spatiotemporal Network

Facial expression recognition refers to the use of computer vision technology to analyze facial expression information, thereby achieving the recognition of human emotions. Existing dynamic expression recognition models struggle to learn key spatiotemporal features related to expressions and accurately model the entire process of expression occurrence. To address these issues, this paper proposes a dynamic expression recognition method based on attention mechanisms and spatiotemporal networks, which introduces spatiotemporal attention mechanisms into deep spatiotemporal networks to focus on key spatiotemporal features during expression changes. Firstly, spatiotemporal features of expression sequences are extracted by introducing spatiotemporal attention modules into multiple 3DCNN blocks. Then, a ConvLSTM network is used to extract global spatiotemporal features. Finally, a label smoothing module is introduced to enhance the model's robustness and generalization ability. Through comparative experiments, ablation studies, and other visualization experiments on the CK+ and Oulu-CASIA datasets, the proposed method is validated to improve the classification performance of expression recognition models.