Facial mask attention network for identity-aware face super-resolution

Face Super-Resolution (FSR) is a crucial research topic in image restoration field, which is a fundamental task for subsequent face applications, such as cross- and low-resolution face recognition. Recently, supported by deep convolutional neural networks, the previous FSR methods have achieved great success in generating high quality face images. However, they mainly focus on improving the visual effects of the images while retaining a challenge of restoring identity information from low-resolution faces. Specifically, some face structure information is discarded, such as the position and the shape of the face components, containing useful identity-related details. To solve this issue, we propose the Facial Mask Attention Network utilizing this information to generate faces of both high resolution and identity fidelity. Furthermore, we present an efficient pixel loss function, MaskPix loss, which selectively emphasizes those significant pixels to focus the model on the face regions with dense identity features. Extensive experiments on popular datasets demonstrate that our restored face images not only have more natural textures and facial details, but also preserve higher identity fidelity compared to the state-of-the-art methods.