An Optically-encoded Loss-predictive Framework for Face Recognition Using Nonlinear Adaptive Margin

Abstract

Recent face recognition strategies using deep neural networks (DNNs) mainly focus on the development of new loss functions and the evolution of network architecture. Due to the large capacity of face datasets, such DNN models usually suffer from a long training time. Motivated by freeform optics design, in this paper we propose a novel paradigm of an optical image encoder, DNN-decoder system for improved face recognition. To make the model learn better from unfamiliar samples, we introduce a covariance loss prediction module attached to the network backbone to dynamically adjust the loss objective. The model defines a nonlinear adaptive margin to measure the angular distance between high-dimensional features and utilizes a PID optimizer to update its parameters, resulting in a faster convergence. Empirical results have shown that the proposed model achieves higher training efficiency on public large training datasets such as WebFace42M, MSIMV2 and CASIA-WebFace, and enjoys state-of-the-art recognition performance on popular evaluation datasets including LFW, MegaFace and IJB-C.