On the Effectiveness of Visible and MWIR-Based Periocular Human Authentication When Wearing Face Masks

Abstract

AbstractIn the COVID-19 (Coronavirus Disease-19) era, everyone is advised to wear a face mask. According to the Centers for Disease Control and Prevention (CDC) published on 19th of April 2021 (Centers for Disease Control and Prevention: Guidance for wearing masks, 2021), masks are most effective when everyone wears one to completely cover vital facial features, i.e., nose and mouth, and to fit correctly against the sides of the face, so that there is no air gap. While the benefits of masks to control the spread of a pandemic or even an endemic are known, many challenges are imposed on state-of-the-art face recognition systems. According to a study published by the National Institute of Standards and Technology (NISTIR 8311), the accuracy of facial recognition algorithms is reduced between 5 and 50% when compared to the accuracy yielded by the same algorithms when the subjects are not wearing face masks. It has been observed that the periocular region is the feature-rich region around the eye, which includes features such as eyelids, eyelashes, eyebrows, tear ducts, eye shape, and skin texture. The report from NIST also states that face images of subjects wearing masks can increase the failure to enroll rate (FER) more frequently than before. In addition, masked face images lower the efficiency of surveillance (unconstrained) face recognition systems, which become even more prone to error due to occlusion, distance, camera quality, outdoors, and low light. In this study, we focus on the effectiveness of periocular-based (combining both eye regions and not just one eye, i.e. left or right) face recognition algorithms when the subjects are wearing face masks under controlled and challenging conditions both for visible and MWIR (mid-wave infrared) band face images. We utilize MILAB-VTF(B), a challenging multi-spectral face dataset composed of thermal and visible videos collected at the University of Georgia (the largest and most comprehensive dual band face dataset to date) (Peri et al (2021) A synthesis-based approach for thermal-to-visible face verification. In: 2021 16th IEEE international conference on automatic face and gesture recognition (FG 2021). IEEE, pp 01–08). We manually crop the faces from the images and use the existing pre-trained face recognition algorithms to perform periocular face recognition. The FR models used in this research study are FaceNet, and VGG-Face. After manually cropping the faces, we perform same-spectral periocular face recognition. FaceNet yields a Rank-1 face identification accuracy of 87.54 and 83.54% in the thermal and visible bands respectively, while VGG-Face yields superior performance with 100 and 99.52% in the thermal and visible bands respectively. Additionally, we also perform same-spectral face recognition experiments (visible-to-visible and thermal-to-thermal) and report the results.KeywordsCovid-19face maskPeriocularBiometric authenticationMILAB-VTF(B)Large-scale dataThermal bandVisible band