注意力引导下的面部动作单元级表情编辑方法——AUA-GAN  

AUA-GAN:facial action unit level expression editing method guided by attention mechanism

作　　者：彭一彤 杨爱英[1,2] PENG Yitong;YANG Aiying(School of Optics and Photonics,Beijing Institute of Technology,Beijing 100081,China;Key Laboratory of Photonics Information Technology,Ministry of Industry and Information Technology(Beijing Institute of Technology),Beijing 100081,China)

机构地区：[1]北京理工大学光电学院,北京100081 [2]信息光子技术工信部重点实验室(北京理工大学),北京100081

出　　处：《计算机应用》2021年第S01期36-42,共7页journal of Computer Applications

基　　金：国家自然科学基金资助项目(61971046)。

摘　　要：为解决图像转换过程中产生的伪影问题,利用生成对抗网络(GAN)生成逼真的人脸表情变化,提出了一种注意力引导下的面部动作单元(AU)级表情编辑方法。首先,在数据预处理部分加入正脸恢复模块,当输入图像的姿态偏转较大时,先经过正脸恢复再进行表情编辑,可以有效提高表情生成质量。其次,生成模块中的生成器和判别器网络内置注意力机制,使图像转换集中在人脸区域,忽略不相干的背景信息。最后,在公开数据库CelebA上训练模型,并选取CK+和CASIA-FaceV5数据库进行图像生成实验。结果表明生成图像与目标图像间的结构相似性(SSIM)为0.804,生成图像的平均表情识别准确率为0.644,重建图像与真实图像间的SSIM为0.951。AUA-GAN可以在较好地保持原有身份信息的前提下,生成清晰准确的人脸表情变化。In order to solve the problem of artifacts in the image conversion process,using Generative Adversarial Network(GAN)to generate realistic facial expression conversion,an attention-guided facial Action Unit(AU)level expression editing method was proposed.Firstly,a front face recovery module was added to the data preprocessing part,which could effectively improve the quality of expression generation.Secondly,both of the generator and discriminator networks had built-in attention mechanism,which made the image transformation focus on the face area and ignore the irrelevant background information.Finally,the model was trained on CelebA database,and the image generation experiments were carried out on CK+and CASIA face V5 databases.The results show that the Structural SIMilarity(SSIM)between the generated images and the target images is 0.804,the average expression recognition accuracy of the generated images is 0.644,and the SSIM between the reconstructed images and the real images is 0.951.The proposed method can obtain high-quality expression conversion images while maintaining the original identity information.

关 键 词：人脸生成 表情编辑 生成对抗网络 动作单元 注意力机制 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]