A Real-Time Deep Learning Framework for Classroom Facial Expression Recognition: Performance Optimization and Model Evaluation

Facial expressions indicate a person’s affective state and can be a significant determinant of cognitive performance. This study proposes a Facial Expression Recognition System (FERS) to detect and analyze students’ real-time emotions, thereby providing teachers with insights. The proposed system was trained and evaluated using eight pretrained models on the CK+ dataset. A comparative analysis indicates that the Xception model achieved the highest accuracy in emotion classification. To improve model performance, the grayscale images in the CK+ dataset were enhanced and used as input to an Xception-based Convolutional Neural Network (CNN), employing 3×3 Conv2D filters with ReLU activation and same-padded layer feature extraction. The model demonstrated excellent performance, achieving an accuracy of 99.34%, a precision of 90%, a recall of 87%, and an F1-score of 88%, confirming the system's reliability and efficiency, applying macro averaging. In conclusion, the proposed Xception-based CNN FERS can accurately recognize students’ emotions, allowing teachers to monitor students' moods in the classroom.