Performance comparison of deep learning based face identification methods for video under adverse conditions

Abstract

Face identification is an important problem in computer vision, which has many application areas. Recently, a number of deep-learning-based face identification and verification methods have been proposed in the literature, which demonstrate remarkable results on large image and video databases. Although the databases used for training and testing deep-learning architectures contain illumination, head pose, and expression variations, they do not reflect the difficult distortions (such as blur and low resolution), which may be encountered when using data from various sources (e.g. surveillance cameras). In this work, our goal is to systematically compare the performance of recent deep-learning-based methods for face identification using video under challenging conditions. We evaluate three deep learning architectures OpenFace, VGGFace2, and ArcFace. The experimental results indicate that even the most successful deep-learning-based face identification methods show poor performance under challenging distortions on the images such as noise, blur and contrast variations. © 2019 IEEE.