Video-based Facial Expression Recognition of Students in Classroom Scenarios Using 3D Convolutional Neural Networks and Multi-Head Self-Attention

The recognition of students' facial expressions in classroom scenarios is an important factor in intelligent teaching evaluation. As video sequences grow longer, video-based facial expression recognition (FER) models based on Recurrent Neural Networks (RNN) and Long Short-Term Memory (LSTM) exhibit high complexity and are prone to issues such as gradient explosion and vanishing, making them difficult to capture long-term dependencies. This paper proposes a model called C3D-SA, which combines 3D Convolutional Neural Networks (CNN) and Muli-Head Self-Attention (MHSA). By introducing MHSA into 3D CNN for spatio-temporal feature extraction from video expression sequences, and emphasizing video frames highly relevant to expression recognition, the model can learn long-term dependencies with lower complexity compared to RNN and LSTM. Moreover, MHSA can extract richer feature information compared to a single-head mechanism. Experimental results on CK+, Oulu-CASIA and AFEW datasets show that C3D-SA achieves higher accuracy compared to 3D CNN and other methods, and can effectively recognize students' facial expressions.