This paper presents EduVision, a content-aware student engagement analysis framework that integrates facial emotion recognition with instructional theme alignment in real classroom environments. A self-collected dataset was recorded from English and Mathematics lectures. Facial expressions were classified using a ResNet18-based deep learning model, and lecture audio was transcribed to map emotional responses to instructional themes. Dimensionality reduction techniques including Principal Component Analysis and Autoencoders were applied, followed by K-Means clustering to identify engagement states. Results show that Autoencoder-based clustering achieved stronger engagement separation, with silhouette scores ranging from 0.61 to 0.65. The analysis further revealed subject-dependent engagement differences, with English explanation segments producing higher positive emotional response than Mathematics sessions. The study demonstrates how aligning emotional analytics with instructional context can provide actionable insight for AI-driven classroom engagement monitoring systems.
Ayesha Jahangir (Tue,) studied this question.