%0 Journal Article
%T 基于生成对抗网络的多角度人脸重构<br>Multi-Poses Face Reconstruction Based on Generative Adversarial Networks
%A 赵志舜
%A 程良伦
%A 黄国恒
%J Computer Science and Application
%P 445-455
%@ 2161-881X
%D 2020
%I Hans Publishing
%R 10.12677/CSA.2020.103046
%X <div style=\"text-align:justify;\">
	<span style=\"font-size:14px;\">由于在不同角度下人脸的差异性很大，多角度的人脸重构仍然是一项具有挑战性的研究课题。现存的方法只能建立两个不同人脸角度之间的映射。若要生成多角度人脸图像，只能通过多个生成器进行多次训练来实现任意两个角度之间的映射，这极大地增加了训练的时间成本。本文提出了一种基于生成对抗网络的多角度人脸重构算法。通过引入对抗损失函数和循环一致性损失函数，本文中的算法仅通过一次训练就能建立任意两个人脸角度之间的映射，这极大地降低了训练的时间成本以及减少了训练难度。在公开数据集Multi-PIE上进行对比实验，结果表明本算法不仅能进行多角度人脸转换，而且在人脸转正之后的人脸识别准确率明显高于CPF、DR-GAN、Light-CNN等算法。<br />
Because of the large discrepancy between human faces at different angles, multi-poses face recon-struction is still a challenging research topic. Existing methods can only establish a mapping of two different face poses. To generate multi-poses face images, multiple generators have to be trained multiple times to achieve the mapping between any two poses, which greatly increases the time cost of training. In this paper, a face reconstruction algorithm based on generative adversarial networks is proposed. By introducing a combination of adversarial loss and cycle consistency loss, the algorithm in this paper can establish a mapping between any two face poses through only one training, which greatly reduces the time cost and the difficulty of training. Comparative experiments on the public dataset Multi-PIE show that this algorithm can not only perform multi-poses face generation, the accuracy of face recognition after face frontalization is significantly higher than CPF, DR-GAN, Light-CNN and other algorithms.</span> 
</div>
%K 人脸重构，人脸转正，生成对抗网络<br>Face Reconstruction
%K Face Frontalization
%K Generative Adversarial Networks
%U http://www.hanspub.org/journal/PaperInformation.aspx?PaperID=34518