The face images were firstly preprocessed as described in Section 4.1. In contrast to SMFRD dataset, RMFRD is imbalanced (5,000 masked faces vs 90,000 nonmasked faces). Therefore, we have applied an over-sampling by cropping some non-masked faces to get an equiva- lent number of cropped and full faces. Next, using the normalized 2D faces, we employ the three pre-trained models (VGG-16, AlexNet and ResNet-50) separately to extract deep features from their last convolutional are (14 × 14 × 512, 13 × 13 × 256, 7 × 7 × 2048) dimensional, respectively.