Neural style transfer generative adversarial network (NST-GAN) for facial expression recognition

With the increasing number of intelligent human-computer systems, more and more research is focusing on human emotion recognition. Facial expressions are an effective modality in emotional recognition, enhancing automatic emotional analysis. Although significant studies have investigated automatic facial expression recognition in the past decades, previous works were mainly produced for controlled environments. Unlike recent pure CNN-based works, we argue that it is practical and feasible to recognize an expression from a facial image. However, the extracted features may capture more identity-related information and are not purely associated with the specific task of expression recognition. To reduce the personal influence of identity-related features by removing identity information from facial images, we propose a neural style transfer generative adversarial network (NST-GAN) in this paper. The objective is to determine the expression information from the input image by removing identity information and transferring it to a synthetic identity. We employ experimental strategies to evaluate the proposed method on three public facial expression databases (CK+, FER-2013, and JAFFE). Extensive experiments prove that our NST-GAN outperforms other methods, setting a new state of the art.