Training Deep Face Recognition for Efficient Inference by Distillation and Mutual Learning

Currently most of deep face recognition algorithms utilize heavy networks to achieve the state-of-the-art performance. In most scenarios, the more challenging task is to achieve the relative high accuracy with low computational cost especially for embedded devices. In this paper, we propose a lightweight network for face recognition using distillation and deep mutual learning In proposed methods a new indicator is designed to monitor the model convergence and an assessment criteria is developed to evaluate the Labeled Faces in the Wild(LFW) dataset. Experiments show that our models work better than networks trained directly and other mobile face recognition solutions.