郭迎春 闫帅帅 刘依

摘要 针对目前表情生成网络中存在的人脸表情失真、不同帧间图像明暗差异明显的问题，提出一种基于递归双对抗网络模型的人脸表情生成框架。首先通过提取深度人脸特征并生成表情特征图，将其作为监督信号，生成人脸表情种子图像;然后使用生成的种子图像和原始目标人脸一起作为输入，生成特征保持图像，作为当前帧的输出，同时该特征保持图像也作为下一帧种子图像生成的输入;最后，将种子图像生成网络和特征保持图像生成网络递归进行下一帧图像的生成，多次递归得到与原始输入表情一致的特征保持人脸表情视频序列。在CK+和MMI数据库上的实验结果表明，提出的方法能够生成清晰自然的人脸表情视频帧，且在目标人脸形状和驱动的表情特征图像有较大形状差异时具有鲁棒性。

关 键 词 生成对抗;深度学习;表情动画;表情生成;双模型

中图分类号 TP391     文献标志码 A

Abstract Aiming at the problem of facial distortion of animated video generated in current facial expression generation network and large differences of light between different frames， a recursive facial expression synthesis framework based on dual network model is proposed. Firstly， deep facial features are extracted to construct facial feature maps， which are used as supervisory signals to generate seed images of facial expressions. Then， the generated seed images together with the original target faces are used as the input information to generate feature-preserving images as the output of the current frame， and also the output feature-preserving image is used as the input of the next seed image generation. Finally the next frame image is generated recursively to synthesize the facial expression sequence which features consistent with the original input expression. Experimental results on CK+ and MMI database show that the proposed method can generate clear and naturai seed image and feature-preserving image， and it is robust when there is a large difference between the shape of the target face and the shape of the driven facial expression image.

Key words GAN; deep neural network; facial animation; expression generation; dual model

0 引言

人脸表情生成是指将人脸表情从源对象迁移到目标对象，新合成的目标对象人脸图像保持不变，但是其表情与源对象一致，这种技术在影视动画、视频游戏、社交娱乐等领域的应用非常广泛[1-2]。随着面部表情迁移技术的发展，虚拟现实、情感互动、人工智能等领域的应用日益广泛，人脸表情生成引起了大量计算机图形学和计算机视觉领域的相关人员的研究兴趣。

过去几十年来研究人员提出了大量的人脸表情生成的方法，现有的方法可以根据像素的操作分为两大类，分别是基于图形学的方法和基于图像的方法[3]。……