Face super resolution based on attention upsampling and gradient

Face Super-Resolution(SR) is a specific domain SR task, which is to reconstruct low-resolution(LR) face images. Recently, many face super-resolution methods based on deep neural networks have sprung up, yet many methods ignore the gradient information of the face image, which is related closely to the restoration of image detail features. At the same time, many super-resolution methods directly use linear interpolation or pixel shuffle and several convolution layers to up-sample the feature maps, caussing some irrelevant pixels will make subsequent detail reconstruction difficult. Considering these issues, in this paper, we propose a face super-resolution method guided by the gradient structure. In particular, we designed a sub-network to generate gradient information from low-resolution images and up-sample the gradient as additional information for the entire network. Unlike other methods based on prior information, such as facial landmarks, facial parsing, face alignment, the gradient information is generated from low-resolution images. At the same time, relying on pixel shuffle, we also designed a novel upsampling module based on channel attention and pixel attention. The results of the experiment show that our network can achieve the sota on several public datasets on PSNR, SSIM, and VIF. The visual result also proves the feasibility and advancement of our network in restoring the detailed structure.