An end-to-end robust feature learning method for face recognition

Abstract

As for deep learning-based face recognition, training a discriminative feature representation is challenging when dealing with noisy-labeled data. This paper introduces a feature learning method robust to such conditions. Our key contributions include an online data filtering algorithm that automatically segregates correctly labeled data from noisy-labeled training data. Additionally, we propose a mechanism called online negative centers sampling (ONCS), which can enlarge the feature space distance between samples within the same class and the centers of different classes. Thus feature learning can be contributed by all the data with ONCS, including the noise-labeled data. We test our method to training an 128-D feature representation on the extreme noisy MS-Celeb-1M dataset, without any preprocess procedures like pre-training dataset or cleaning dataset. The result demonstrates an accuracy of 99.33% on LFW test set with a single model and without the preprocessing of landmark-based alignment close to the result by the clean data.