Transfer Learning for Robust Masked Face Recognition
Omar Adel Muhi, Mariem Farhat, Mondher Frikha · 2022
Over the past three years, it has seen many outbreaks of different coronavirus diseases around the world. One of the ways of transmission of COVID 19 is air transport. This transmission occurs when humans breathe in droplets released by an infected person by breathing, speaking, coughing or sneezing. The World Health Organization (WHO) has given orders to wear a face mask in the public places.Despite the impressive results achieved by deep learning methods in face recognition, the performance of these methods deteriorates when wearing a mask. The issue of masked face recognition is attracting more attention.In this work, a simple and effective method is proposed to deal with the recognition of people who wear a mask. We divide the problem into three phases: feature detection, feature extraction and recognition. The first stage of our model with MediaPipe Face Mesh automatically produces a segmentation of the masked area as feature detection and several points for cropping the area of interest. Then, the second stage extracts the features gained using Resnet50. The selected network, based on ResNet-50, is modified so that the third stage performs the classification process with Soft Max as an redaction and activation function.To train our system, we used the Labeled Faces in the Wild (LFW). From the original facial recognition data set, a masked copy is generated using data augmentation, and the two data sets are combined during the training process. Our model achieved a test accuracy of 98%.