Effective training of convolutional neural networks for face-based gender and age prediction

作者：

Highlights：

• Distribution label encoding as an optimal way to represent age in a CNN.

• Role of pretraining in selection of the CNN architecture and the training strategy.

• Advantages of mono-task and multi-task trainings.

• State-of-the-art scores on 3 popular benchmarks: LFW, MORPH-II and FG-NET.

• 1st place in the ChaLearn Apparent Age Estimation Challenge 2016.

摘要

•Distribution label encoding as an optimal way to represent age in a CNN.•Role of pretraining in selection of the CNN architecture and the training strategy.•Advantages of mono-task and multi-task trainings.•State-of-the-art scores on 3 popular benchmarks: LFW, MORPH-II and FG-NET.•1st place in the ChaLearn Apparent Age Estimation Challenge 2016.

论文关键词：Gender recognition,Age estimation,Convolutional neural network,Soft biometrics,Deep learning

论文评审过程：Received 23 December 2016, Revised 2 May 2017, Accepted 25 June 2017, Available online 27 June 2017, Version of Record 4 July 2017.

论文官网地址：https://doi.org/10.1016/j.patcog.2017.06.031