<p>Dynamic Facial Expression Recognition (DFER) faces significant challenges in in-the-wild video sequences due to the complexities of capturing spatial and temporal features. To address this challenge, we propose a Dynamic Attention and Multi-Scale Temporal Former (DAMT-Former), which integrates the Spatial Dynamic Attention Former (SDA-Former) and the Temporal Multi-Scale Fusion Former (TMF-Former) to enhance facial expression recognition performance in videos. Specifically, the SDA-Former module dynamically adjusts the spatial feature weights of each frame by calculating the magnitude of facial muscle changes between adjacent frames, thereby suppressing irrelevant information and improving the learning capability of spatial features. The TMF-Former module captures long- and short-term temporal dependencies between frames with significant facial expression changes through a cross-frame attention mechanism and integrates it with global temporal context. The multi-scale mechanism of the TMF-Former further enhances the model’s ability to model temporal features at different time scales. Extensive experiments and ablation studies demonstrate that DAMT-Former consistently improves performance on video-based facial expression recognition tasks and achieves performance that is comparable to or slightly better than recent state-of-the-art methods on the DFEW, FERV39k, and AFEW datasets. Additionally, the visualization of feature distributions shows that this approach effectively learns more discriminative facial features.</p>