Asymmetric Pyramid Based Super Resolution from Very Low Resolution Face Image

摘要

Most of the existing one-step upsampling super-resolution (SR) methods could not clearly reconstruct a higher resolution image from a very low-resolution image because there is not enough supervision information to be available. Inspired by the laplacian pyramid, we propose a novel Asymmetric and Progressive Face Super-Resolution Network (APFSRNet) to progressively reconstruct a super-resolution face image from a very low-resolution face image. To further improve the accuracy of the reconstruction, we use the densely connected layers to deepen our network which also alleviate the vanishing-gradient problem. We use the entire face image to train our network instead of using face image patches to maintain the global structure of the face image. Furthermore, we employ structural similarity index (SSIM) as a part of loss function to satisfy human observation. Our extensive experiments demonstrate the effectiveness of the proposed model qualitatively and quantitatively.