Online Learning Facial Expression Detection using Simplified AlexNet Deep Learning Architecture: Image Data Samples Comparison Experiment

Online learning is one of the methods for students to interact with their teachers. However, interactivity aspects in online education should be evaluated. The method to automate the detection of students' enthusiasm when participating in online learning is still not much found. In this paper, we define the facial expressions that could show when students participate in online learning: bored, confused, and excited/delighted. We built a classifier system for these online learning face expressions. We collected a small dataset (for around 480 images) containing face images with these three facial expressions to be trained for the deep learning model. We proposed simplified AlexNet architecture as a deep learning model for a relatively small dataset, that is, AlexNet architecture with lower parameters (such as filters and the number of neurons in the fully connected layer). Image preprocessing methods such as resizing and augmentation are applied for the experiment. We also compare the model performance, given a distinct (three kinds) number of training data (480 samples, 1,500 samples, and 3,000 samples). When our proposed model is trained with these three kinds of training samples amount, the model performance consecutively has an average accuracy of 78.69%, 84.53%, and 90.80% in 5-Fold Cross Validation. Using ANOVA Test, we get a p-value of 0.00073, which is statistically significant to reject the null hypothesis that training samples do not affect our proposed model's accuracy.