AEPL: Adaptive empirical prototype learning with dynamic margins for deep face recognition

Weijia Fan, Yan Huang, Zhixiang Cai, Chunsong Chen, Yanxi Liu, Jiajun Wen, Xi Jia, Linlin Shen, Jiancan Zhou, Qiufu Li

Abstract: Prototype learning is widely used in face recognition, which commonly takes as the prototypes the row vectors of coefficient matrix in the last linear layer of the feature extraction model for each class. During the model training, when the prototypes are updated using the gradients of facial sample features, they are prone to being pulled away from the ideal class center by the features of hard facial samples, resulting in a decrease in overall model performance. In this paper, we define the prototypes as the expectations of sample features in each class and design empirical prototypes using the features of existing samples in the dataset. We then devise a adaptively updating strategy to these empirical prototypes during the model training using the similarity between the sample features and the empirical prototypes, and propose an adaptive empirical prototype learning method (AEPL), which utilizes an adaptive margin parameter with respect to sample features. Different from the previous prototype learning methods, AEPL assigns larger margins to the normal samples and smaller margins to the hard samples, allowing the learned empirical prototypes to better reflect the ideal class centers that are dominated by features of the normal samples and finally pull the hard samples toward the empirical prototypes through the model training. Extensive experiments on MFR, IJB-C, LFW, CFP-FP, AgeDB, and MegaFace demonstrate the effectiveness of AEPL.

External IDs:dblp:journals/paa/FanHCCLWJSZL26