基于条件循环生成对抗网络的人脸姿态重建

Face pose reconstruction based on conditional cycle generative adversarial networks

人脸姿态重建对于解决由于人脸姿态导致的人脸识别率降低的问题有重要意义,由于自遮挡,缺少大部分的人脸特征,重建正面人脸存在很大困难。近年来运用生成对抗网络的图像生成方法得到学界深入的研究,受生成对抗网络在人脸肤色,头发等属性变换等方面研究工作的启发,将人脸偏转角度作为人脸的一种全局姿态属性进行基于生成对抗机制的互换训练,提出了一种基于条件循环生成对抗网络(CC-GAN)的姿态重建方法,在FERET数据库上生成的人脸图像结构相似度达到了0.75,峰值信噪比达到了23.27 dB,均优于当前的其他方法,取得了良好的人脸重建效果。

Face pose reconstruction is of significant importance for solving the problem of face recognition rate reduction caused by face pose.Due to self-occlusion and lack of most face features,it is very difficult to reconstruct the front face.In recent years,image generation methods based on generative adversarial network have received in-depth researches.Inspired by the research work of generative adversarial network in the face attribute(color,hair,mouth,etc.)exchanges,face deflection angle is used as a global pose attribute for face exchange training based on generative countermeasure mechanism,thus a face pose reconstruction method based on conditional cycle generative adversarial network(CC-GAN)is proposed.The similarity of face image structure generated on the FERET database is 0.75,and the peak signal-to-noise ratio(PSNR)is 23.27 dB,which are both superior to other current methods in face reconstruction performance.