Selfies for mobile biometrics: Sample quality in unconstrained environments

Abstract

Taking a ‘selfie’ using a mobile device has become a natural gesture in everyday life. This simple action has many similarities to face authentication on a smartphone: positioning the camera, adjusting the pose, choosing the right background and looking for the best lighting conditions. In the context of face authentication, most of the standardised processes and best practice for image quality is mainly focused on passport images and only recently has the attention of research moved to mobile devices. There is a lack of an agile methodology that adapts the characteristics of facial images taken on smartphone cameras in an unconstrained environment. The main objective of our study is to improve the performances of facial verification systems when implemented on smartphones. We asked 53 participants to take a minimum of 150 ‘selfies’ suitable for biometric verification on an Android smartphone. Images were considered from constrained and unconstrained environments, where users took images both in indoor and outdoor locations, simulating real-life scenarios. We subsequently calculated the quality metrics for each image. To understand how each quality metric affected the authentication outcome, we obtained biometric scores from the comparison of each image to a range of images. Our results describe how each quality metric is affected by the environment variations and user pose using the biometric scores obtained. Our study is a contribution to improve the performance and the adaptability of face verification systems to any environmental conditions, applications and devices.