Masked Face Recognition Using Convolutional Neural Networks and Similarity Analysis

Nowadays, human face recognition systems have been widely used in different applications in which identity recognition is needed. The performance of current face recognition algorithms is negatively affected by occlusions, such as facial masks and various human poses. To address these challenges, we re-trained a modified version of the VGG19 deep learning model on masked and unmasked images of 62 identities to design a feature extractor that extracts deep features from the non-occluded areas of the face. This feature extractor is combined with our proposed similarity analysis network that is trained on our dataset to automatically judge whether the masked and unmasked images correspond to the same or different identities. Our final approach consists of a feature extractor from a fine-tuned VGG19 and a similarity model. It achieved an accuracy of 80 to 85 percent in recognizing the identity of test masked images with different poses.