Dynamic facial expression recognition based on attention mechanism

Abstract In the wild, dynamic facial emotion recognition is a highly challenging task. Traditional approaches often focus on extracting discriminative features or preprocessing data to remove noisy frames. The former overlooks differences between keyframes and noise frames, while the latter can be complex and less robust. To address this issue, we propose a spatiotemporal feature extraction network based on an attention mechanism. In the spatial feature extraction stage, our method incorporates prior knowledge through an attention mechanism, allowing the model to precisely select and focus on key features. Additionally, incorporating prior knowledge reduces unnecessary search space, leading to faster convergence to optimal solutions. To suppress the impact of irrelevant frames, our study introduces a dynamic-aware temporal feature extraction mechanism, enabling the network to accurately capture key frames in video sequences, thereby enhancing model performance. Finally, at the decision stage, we design a fragment-based feature fusion filter to mitigate the influence of irrelevant frames on sequence classification. Our method achieves accuracy rates of 54.18% on DFEW, 42.37% on FERV39k, and 75.55% on CAER tests.