Towards a Masked Face Recognition Algorithm: A Novel Rule Based Hybrid Algorithm

• Published in: 2021 6th South-East Europe Design Automation, Computer Engineering, Computer Networks and Social Media Conference (SEEDA-CECNSM) pp. 1 - 6
• Main Authors: Alexiou, Michail, Ktistakis, Iosif Papadakis, Goodman, Garrett
• Format: Conference Proceeding
• Language: English
• Published: IEEE 24-09-2021
• Subjects: COVID-19; Face Masks; Face recognition; Healthcare; Iris; Lighting; Machine learning algorithms; Prediction algorithms; Smart cities; Social networking (online)
• DOI: 10.1109/SEEDA-CECNSM53056.2021.9566244

The recent SARS-CoV-2 virus global spread and the Covid-19 pandemic that resulted from that has increased the focus on hygienic and contactless safety measures. The wide use of masks, that are essential for the reduction of the spread of the virus, has been declared mandatory by several institutions and countries. Given the circumstances, masked facial recognition has gained an increased attention by the research community. Mask identification for healthcare reasons, face recognition and identification for public safety in a smart city environment and security reasons proved to be a challenging problem with most algorithms having limitations. The vast amount of data, of people with masks, that has been generated in the recent months will be helpful in tackling some of the major issues. The main issues most methodologies face are occlusion, illumination, non-frontal characteristics and pose variation. This paper focuses on the preliminary results of a novel rule based hybrid masked face recognition algorithm. We use the MaskedFace-net dataset and we detect the covered face using the Viola-Jones algorithm. Through the use of Statistical Region Merging (SRM) we detect the ocular region in the cropped face image, and we achieve eye detection for our preprocessing. Through this process we manage to get a clearer and sharper version of the original input eye image. Finally, we generate the attributed graph of the detected facial features, whose labeled arcs represent the computed distance rations between them. This algorithm will act as input into a machine learning prediction model moving forward.