A feature learning approach for face recognition with robustness to noisy label based on top-N prediction

Abstract

Collecting a vast amount of face data with identity labels to train a convolutional neural network is an effective mean to learn a discriminative feature representation for face recognition. However, the datasets with larger scale often contain more noisy labels, that directly affects the ultimate performance of the learned model. This paper proposes an end-to-end feature learning method with robustness to noisy label. First, a data filtering method is proposed to automatically online filter out the data with false label, by checking the consistency between the annotated label and the results of top-N prediction. Then the loss functions of softmax and center loss are simply revised to only supervise the reserved feature. Finally, we use MS-Celeb-1M dataset, which contains massive noisy labels, to train a 128-D feature representation without any pre-train or data pre-clean. A single learned model gets an accuracy of 99.43% on LFW test set, that is very close to the model trained using the clean data.