Landmark perturbation-based data augmentation for unconstrained face recognition

作者：

Highlights：

• Landmark perturbation-based data augmentation method is able to generate different kinds of artificial face images automatically.

• The trained DCNN model using landmark perturbation-based data augmentation method is robust to misalignment.

• The proposed data augmentation method improves face recognition rates, meanwhile it provides faster convergence of DCNN training at early stage.

摘要

Highlights•Landmark perturbation-based data augmentation method is able to generate different kinds of artificial face images automatically.•The trained DCNN model using landmark perturbation-based data augmentation method is robust to misalignment.•The proposed data augmentation method improves face recognition rates, meanwhile it provides faster convergence of DCNN training at early stage.

论文关键词：Feature representation,Face recognition,Landmark perturbation,Misalignment,Deep convolutional neural networks

论文评审过程：Received 8 September 2015, Revised 31 March 2016, Accepted 31 March 2016, Available online 1 April 2016, Version of Record 29 September 2016.

论文官网地址：https://doi.org/10.1016/j.image.2016.03.011