Academic Emotion Classification Using FER: A Systematic Review

Table 6

Selected publications that used academic emotion dataset.

Reference	Emotion classifier	Type of emotions	Feature extraction	Results	Limitation	Real time
[97]	Long-term recurrent convolutional network	Boredom, confusion, engagement, and frustration	N/A	Top-1 accuracy: 94.6%	Low illumination and lack of frontal pose	No
[35]	Deep belief network (DBN)	Two-level engagement and three-level engagement	Viola-Jones algorithm	Accuracy: (i) Two-level: 90.89%; (ii) Three-level: 87.25%	Unknown direct correlation between engagement and actual task performance	No
[46]	CNN	Confusion, distraction, enjoyment, fatigue, neutral	VGG16	Accuracy: 91.60%	Limited sample number	No
[47]	BERN	Boredom, confusion, engagement, and frustration	OpenFace	Top-1 accuracy: 60% for four classification model	Requires a large amount of training data and a long training time	Yes
[52]	LSTM	Boredom and frustration	MTCNN library—face detection & cropping	Accuracy compared to the EmotioNet model: (i) Boredom: +16.26%; (ii) Frustration -2.42%	(i) Decreased accuracy for frustration emotion; (ii) Involved negative emotions only	No
[61]	CNN	Boredom, confusion, engagement, and frustration	CNN+pose estimator	Accuracy: 53.4%	Show a low recognition rate for the frustration when using the DAiSEE dataset	Yes
[64]	Deep facial spatiotemporal network (DFSTN)	Engagement level: very low, low, high, and very high	MTCNN	Accuracy: 58.84%	(i) Low detection accuracy; (ii) Data deficiencies and data imbalances	No