Rethinking Pixel-Wise Loss for Face Super-Resolution

Face super-resolution has become much attractive in terms of its applications, i.e., video surveillance, and face identification system. However, it is difficult to reconstruct super-resolved images from low-resolution images due to the lack of information in missing pixels. To achieve it, most state-of-the-art methods have used deep neural networks to compensate for missing parts in the reconstruction process by optimizing pixel-wise loss functions. In these methods, various basic pixel-wise loss functions were utilized in image-level and feature-level, but there did not exist the analysis of pixel-wise functions. In this paper, we analyze pixel-wise loss functions in different image domains and feature-level to explore the effects of the loss function in face super-resolution. Furthermore, we provide extensive experiments on CelebA to rethink existing pixel-wise loss and make the best combination of existing loss functions.