图像生成和深度度量学习的身份感知面部表情识别方法

Facial expression recognition (FER) is a challenging task because the external environment and identity characteristics could affect the classification results directly. To settle down the above-mentioned challenges, this paper proposed an identity-aware facial expression recognition method which combined images synthesis techniques and deep metric learning, and made facial images features compared then classified by creating expression groups under the same identity in FER task. There are three parts in our method. The first part is a generative adversarial network, which aims to learn expression information and synthesis the expression groups. the second part is the feature extraction network, which transforms the image into feature vectors that could be used for metric learning. The third part is Mahalanobis metric learning network that could compare and classify a pair of feature values effectively. The average accuracy of proposed method reached 98.653 2% and 99.824 8% on two well-known FER dataset named CK+ and Oulu-CASIA, with more than 10% higher than the method proposed currently. By comparing with several state-of-the-art methods, the experimental results confirmed that the proposed-method was effective and progressive in FER task.