摘要
Facial expression manipulation has two objectives: 1) generating an image with target expression; 2) preserving the identity information of the original image as much as possible. Recently, Generative Adversarial Networks (GANs) have shown the abilities for fine-grained facial expression manipulation. However, current methods are still prone to generate images with poor quality. In this work, we propose a U-Net based generator with multi-attention gate for facial expression manipulation. The multi-level attention mechanism is helpful to manipulate relevant regions and preserve identity features, thus improving the editing ability. Furthermore, we adopt self-attention block to replace direct skip-connection to get long-range dependency in images. To suppress artifacts in generated images, we add a discriminator based loss function in the training process. Extensive experiments on both quantitative and qualitative evaluation show that our proposed method achieves better performance for facial expression manipulation.
