Face Recognition Using Convolutional Neural Network Architectures on Mask-Occluded Face Images

In epidemic situations such as the novel coronavirus disease (COVID-19) pandemic that spreads through physical contact, security and presence systems that previously used fingerprints-based or were contact-based are no longer safe for users. Compared to other popular biometrics such as fingerprints, irises, palms, and veins, the face has much better potential to recognize identity in a nonintrusive manner. Therefore, this study will employ two convolutional neural network (CNN) architectures, LeNet-5 and MobileNetV2, for face recognition on mask-occluded face images. Data were taken from 12 subjects face-to-face were preprocessed by cropping, artificial mask augmentation, resizing, and image augmentation. The model was trained with the configured hyperparameter for 50 epochs with a 60:40 data split. Model testing was performed using image data without augmentation wearing a mask. The test results are measured with classification accuracy for 12 classes. The highest testing accuracy on LeNet-5 models is 98.15%, with
$64\times 64$
input size and 64 batch size. Meanwhile, the highest testing accuracy for MobileNetV2 is 97.22% with input size
$96\times 96$
, batch size 16, and the weight of the MobileNetV2 model initialized with ImageNet
$96\times 96$
.