维度情感模型下的表情图像生成及应用

Synthesis of the expression image and its application under the dimentional emotion model

为了解决基于深度学习的人脸表情识别所需训练数据包含表情类别有限且训练数据规模不均衡的问题,提出了Arousal-Valence维度情感空间中基于生成对抗网络的表情图像生成方法AV-GAN,用于生成更多样且均衡的表情识别训练数据。该方法使用标记分布表示表情图像,通过引入身份控制和表情控制模块,以及对抗学习方法实现在Arousal-Valence空间中随机采样和生成表情图像。在Oulu-CASIA数据库上的评估实验显示,使用本文方法对训练数据进行数据增强比使用原训练数据的表情识别准确率可提升6.5%,证明了该方法能有效地提升非均衡训练数据下的表情识别准确率。

In order to solve the problem that the training data of deep learning based facial expression recognition methods usually cover a limited part of the expression space and have an imbalanced distribution,we propose AV-GAN,a facial expression image synthesis method in Arousal-Valence dimensional emotion space,based on the generative adversarial network,to generate more diverse and balanced facial expression training data.The method uses label distribution to represent the expression for the face image,and employs an identity control module,an expression control module,and adversarial learning to realize the random sampling and generation of expression images in Arousal-Valence space.Evaluations on Oulu-CASIA database show that the accuracy of the recognition of the facial expression using the proposed method to augment training data is increased by 6.5%,compared with that using the original training data.It is proved that the proposed method can effectively improve the facial expression recognition accuracy under imbalanced training data.