针对非正面姿态下人脸识别率低下的问题,提出一种基于双路循环生成对抗网络的多姿态人脸识别方法。该网络由人脸转正及人脸旋转两部分组成。人脸转正部分完成侧面人脸向正面人脸的转化,实现多对一的姿态类别映射;人脸旋转部分完成对正面人脸身份特征的提取及指定姿态人脸的生成,实现一对多的姿态类别映射。训练过程中利用两条循环路径将人脸转正及人脸旋转过程结合,一路完成人脸侧面至正面再至侧面的循环转化,另一路则完成人脸正面至侧面再至正面的循环转化,促使两部分内容相互利用、约束,提高对侧面人脸的识别率。为了加快网络的收敛速度,降低训练难度,训练过程分为先局部后整体两个不同的阶段进行。在人脸数据集Multi-PIE及CFP上的实验结果表明,该方法能够有效提高对侧面人脸的识别率。

This study proposes a multi-pose face recognition method with a two-cycle generative adversarial network to address low face recognition accuracy of non-frontal poses. The network consists of two aspects: face frontalization and face rotation. The face frontalization aspect converts profile faces to frontal faces and implements many-to-one pose category mapping. The face rotation aspect converts frontal faces to profile faces with specified poses, extracts the identity features of the frontal faces, and implements one-to-many pose category mapping. To further improve the face recognition of profile poses, two cyclic paths are used to combine the face frontalization and face rotation processes. One path is used for the cyclic conversion of profile faces to frontal faces and then to profile faces, and the other path is used for the cyclic conversion of frontal faces to profile faces and then to frontal faces. To reduce the difficulty in the training process and speed up the convergence of the network, the training process will be performed in two different stages: partial and complete training. Experiment results on Multi-PIE and CFP show that this method can effectively improve the recognition accuracy of profile poses.