Extending Earth Mover's Distance to Occluded Face Verification

Facial recognition is one of the most used biometric clues to identify and verify individuals regarding access to secure facilities or devices, for law enforcement purposes, or to locate missing persons. This is due to its non-intrusive nature and high distinguishable power for identity authentication. Deep learning systems have been shown to be the go-to choice to extract distinguishable features from faces. However, a common real world scenario challenge occurs when occlusions (eg. sunglasses, masked, scarf) in the input to such systems decrease their overall performance. To address this issue, the scientific community proposed approaches and competitions, resulting in different solutions to this problem. We focus on extending the Earth Mover's Distance (EMD) to the occluded face recognition prob-lem by evaluating its potential with state-of-the-art backbones on verification tasks by using the OCFR-2022 benchmark. We confirm the applicability of fusing the cosine similarity and EMD distance scores to enhance conservative decision-making process. We draw this conclusion by considering lowering the False Match Rate operation points on verification set.

I. Introduction

Face recognition is a popular method for authenticating and identifying users, such as for access to secure facilities or devices, for law enforcement purposes, or to locate missing persons. A closely related technology is face verification, which is the process of verifying that a person is who they claim to be. It is a one-to-one matching problem, where the system is trying to match the face in the image to a specific person's identity. Face recognition or verification relies on image processing to extract features from faces. These features are then used as input to pattern recognition methods that can identify and match faces. Increasingly, these pattern recognition methods are based on machine learning, such as deep learning networks. Deep learning has been shown to be effective in extracting information from facial images. Trained on large data sets of facial images, deep learning models can learn to identify and extract a wide variety of features from faces, such as the shape of the face, the eyes, the nose, the mouth, and the eyebrows.