Minimum margin loss for deep face recognition

作者：

Highlights：

• Training with unbalanced data leads to a performance drop on face recognition.

• Penalising the overclose classes can alleviate the unbalancedness of data.

• Setting a minimum margin for all the pairs of class centre improves the performance.

• With the proposed loss function, we achieved the state-of-the-art performance.

摘要

•Training with unbalanced data leads to a performance drop on face recognition.•Penalising the overclose classes can alleviate the unbalancedness of data.•Setting a minimum margin for all the pairs of class centre improves the performance.•With the proposed loss function, we achieved the state-of-the-art performance.

论文关键词：Deep learning,Convolutional neural networks,Face recognition,Minimum margin loss

论文评审过程：Received 18 September 2018, Revised 29 May 2019, Accepted 17 August 2019, Available online 17 August 2019, Version of Record 26 August 2019.

论文官网地址：https://doi.org/10.1016/j.patcog.2019.107012