An Empirical Comparative Analysis of Africans with Asians Using DCNN Facial Biometric Models

Abstract

Recently, the problem of racial bias in facial biometric systems has generated considerable attention from the media and biometric community. Many investigative studies have been published on estimating the bias between Caucasians and Asians, Caucasians and Africans, and other racial comparisons. These studies have reported inferior performances of both Asians and Africans when compared to other races. However, very few studies have highlighted the comparative differences in performance as a function of race between Africans and Asians. More so, those previous studies were mainly concentrated on a single aspect of facial biometrics and were usually conducted with images potentially captured with multiple camera sensors, thereby compounding their findings. This paper presents a comparative racial bias study of Asians with Africans on various facial biometric tasks. The images used were captured with the same camera sensor and under controlled conditions. We examine the performances of many DCNN-based models on face detection, facial landmark detection, quality assessment, verification, and identification. The results suggested higher performance on the Asians compared to the Africans by most algorithms under the same imaging and testing conditions.