BCGAN: Facial Expression Synthesis by Bottleneck-Layered Conditional Generative Adversarial Networks

摘要

Facial expression synthesis is widely applied to emotion prediction and face recognition for human-computer interaction. This task is challenging because it is difficult to reconstruct realistic and accurate facial expressions. Early deep learning methods focus only on pixel-level manipulation and are not suitable for generating realistic facial expressions. In this paper, we propose a bottleneck-layered conditional generative adversarial networks (BCGAN) for more realistic and accurate facial expression synthesis. BCGAN adopts a bottleneck layer that uses channel-wise concatenation in the generator to train with meaningful features only. In addition, a dense connection that links all bottleneck layers is added to generate an image which preserves the facial details of the original image. Both quantitative and qualitative evaluations were performed using the Radboud Faces Database (RaFD). Experimental results showed that BCGAN had 2% higher classification accuracy (98.7%) on the generated images as well as faster training speed compared to state-of-the-art approach.