Table 4.

Accuracy, precision, recall and F1-score achieved by the methods tested in the four scenarios and compared to the results achieved in the original datasets without masks and eye occlusions

Model	Dataset	Metric	Original	Scenario1	Scenario2	Scenario3	Scenario4
Residual masking	FER2013	accuracy	74.14%	32.97%	62.03%	54.50%	68.51%
network [18]		precision	74.36%	42.78%	64.12%	54.71%	68.61%
		recall	74.14%	32.97%	62.03%	54.50%	68.51%
		F1-score	74.25%	37.24%	63.06%	54.60%	68.56%
FER CNNs [7]	FER2013	accuracy	62.90%	32.10%	54.20%	45.58%	56.80%
		precision	63.10%	33.24%	59.09%	48.09%	57.44%
		recall	62.90%	32.10%	54.20%	45.58%	56.80%
		F1-score	63.00%	32.66%	56.54%	46.80%	57.12%
Amend-Representation	RAF-DB	accuracy	90.42%	45.45%	82.30%	74.25%	84.32%
Module [27]		precision	90.26%	56.16%	81.75%	74.25%	84.03%
		recall	90.42%	45.45%	82.30%	74.56%	84.32%
		F1-score	90.34%	50.24%	82.02%	74.40%	84.17%