I. Introduction
Face identification (FI) is one of the most intensively investigated topics for researchers in the fields of pattern recognition and computer vision. In the past 20 years, we have witnessed emerging FI methods, including subspace analysis methods [1]–[3], regression-based methods [4]–[6], and convolutional neural network (CNN)-based methods [7]–[13]. The CNN-based methods have recently captured much attention [14]–[16], and many successful deep CNNs are proposed for face recognition, such as FaceNet [7], CosFace [8], and ArcFace [9]–[11]. However, deep learning methods generally require a large amount of training data and high computational power. Besides, the unclear theoretical understanding of deep learning models makes it difficult to determine the optimal architecture and optimization algorithm. The subspace analysis methods can learn discriminative information in data, but they are incapable to well deal with the complex variations in images such as facial occlusions [3], which are commonplace in epidemic.