Semantic attention adaptation network for face super-resolution

Face super-resolution (Face SR) is a sub-domain of SR that reconstructs high-resolution face images from low-resolution ones. The prior knowledge of face is widely used for recovering more realistic facial details, which will increase the complexity of the network and introduce additional knowledge extraction procession both in the training and evaluating stage. To address the above issues, we propose to combine face semantic prior extraction and face SR with the attention adaptation model and design a Semantic Attention Adaptation Network (SAAN) for face SR. Specifically, we train the face semantic parsing network and face SR network jointly, by adopting the semantic attention adaptation (SAA) model to transfer the ability of extracting face prior knowledge to the SR network. Then our SR network can work independently in the testing stage without using the prior knowledge extraction network. To generate realistic face images, we also utilize GAN loss to enrich the texture with more details (i.e. SAAN-G). Extensive experiments on the benchmark dataset illustrate that our SAAN and SAAN-G improve the state-of-the-art both on quality and efficiency.