Facial Expression Recognition in Online Course Using Light-Weight Vision Transformer via Knowledge Distillation

It is essential for teachers to master students’ current learning status in online courses so that they can adjust teaching mode and rhythm. To accurately describe students’ emotion changing during studying online course, we propose a framework that combines a light-weight facial expression recognition(FER) classification model-MobileViT, which has been improved by modifying the fusion block and enhanced by introducing knowledge distillation(KD), with an online course platform. First of all, the face detecting, tracking and clustering are applied to extract face sequence of each student. Then, an improved MobileViT is used to extract emotional features in each frame of online course for classification and prediction. Students’ facial images are collected via the camera of their devices, analyzed using MobileViT, and classified into 7 basic emotions. Our improved MobileViT is efficient because it can be processed in real time on each student’s mobile device which does not require to send their facial videos to teacher’s PC or remote server. Finally, the proposed improved MobileViT is tested on public face datasets RAFD, RAF-DB and FER2013 comparing with some mainstream models. Experimental results indicate that our model has competitive results and better efficiency than others.