Long Short-Term Perception Network for Dynamic Facial Expression Recognition

Dynamic facial expression recognition (DFER) presents a difficult challenge, and antecedent methodologies leveraging convolutional neural networks (CNNs), recurrent neural networks (RNNs), or Transformers focus on extracting either long-term temporal information or short-term temporal information from facial videos. Unlike prevailing approaches, we design a novel framework named long short-term perception network (LSTPNet). It can easily perceive aforementioned dual temporal cues and bestow notable advantages upon the DFER task. To be specific, a temporal channel excitation (TCE) module is proposed, building upon the previous outstanding efficient channel attention (ECA) module. This extension serves to imbue the backbone network with temporal attention capabilities, thereby facilitating the acquisition of more enriched temporal features. Furthermore, we design a long short-term temporal Transformer (LSTformer) which can capture both short-term and long-term temporal information with efficacy. The empirical findings, as showcased across three benchmark datasets, unequivocally demonstrate the state-of-the-art performance of LSTPNet.