Assessment of Face Recognition Algorithms over Periocular and Binocular Areas of the Human Face to Check Authentication Accuracy in the Event of Occlusion

Facial recognition, although less accurate than iris recognition, offers greater hygiene compared to fingerprint and handprint recognition. It is also less affected by ocular pathologies than iris recognition and requires less cooperation than both fingerprint and iris recognition. As a result, facial recognition is widely preferred for identifying subjects in a crowd. However, faces can be occluded for various reasons (wearing sunglasses, facial masks, headgear, and clothes, people using their hands or other objects to cover their mouth and nose), prompting our exploration of binocular and periocular authentication systems and their associated deep learning models. Assessing the resilience of such models to occlusion is a challenge. In this chapter, we evaluate the performance of multiple pretrained convolutional neural network (CNN) image representation models on facial, binocular, and periocular data collected from 100 subjects. These subjects are a part of the MILAB-VTF(B) dataset, and we used two images per subject in both the gallery and probe sets. As a part of our study and to establish a baseline, we first crop the binocular and periocular areas from the original face images using the Viola-Jones algorithm. Then, we perform a set of identification experiments by frontalizing the face with the multitask cascaded convolutional neural network (MTCNN is a deep learning-based face detection algorithm that is also capable of identifying facial landmarks such as the eyes, nose, and mouth). We calculate the cosine distance between the model embeddings of probe images to the gallery images and select the nearest neighbor. We perform similar experiments with generated templates, blocking and isolating areas of interest. Of the nine models tested, ArcFace had the highest accuracy on face versus face (100%), face versus covered binocular (95%), and face versus covered right periocular (100%) matching scenarios.