Face Image Quality - A Comparative Study of Motion Blur Detection Algorithms

Abstract

Motion blur degrades face image quality and impairs recognition accuracy. This paper evaluates five face image quality assessment (FIQA) algorithms for motion blur detection, focusing on accuracy and demographic fairness. Experiments on the EDAMB and MST-E datasets employed Kullback–Leibler (KL) divergence to compare algorithm scores against expert consensus, partial area under the curve (pAUC) from error-versus-discard curves to report prediction of recognition performance, and the Gini coefficient to assess fairness. Densenet169 had the lowest KL divergence, while CNN-R showed the best predicting performance, achieving the lowest pAUC. Fusing CNN-R with Densenet161 further reduced the pAUC by 1.3%. The fairness analysis found that the Fourier Transform and CNN-R methods were the most fair, whereas Laplace was the least fair.