Facial Expression Recognition: Disentangling Expression Based on Self-attention Conditional Generative Adversarial Nets

摘要

The accuracy of facial expression recognition is greatly impacted by individual attributes. To address this problem, we propose a Disentangle Expressions based on Self-Attention Conditional Generative Adversarial Nets method, where facial expression recognition takes by two steps. The first step constructed a generative model to generate the corresponding neutral face image and disentangle expression features. The second step trained the classifier with preserved disentangled expression features. A self-attention layer is used to learn correlations among different facial motion units. Inspired by the relativistic GAN [1], we use the discriminator to predict the relative realness of the generated images and provide strong supervision for more details recovery. The results from extensive experiments on three public facial expression datasets (CK+ , MMI, Oulu-CASIA) proved that our method is more effective than the known state-of-the-art methods in recognition accuracy.