Learning Spatial Attention for Face Super-Resolution

General image super-resolution techniques have difficulties in recovering detailed face structures when applying to low resolution images. Recent deep learning based methods tailored for images achieved improved performance by jointly trained with additional task such as parsing and landmark prediction. However, multi-task requires extra manually labeled data. Besides, most of the existing works can only generate relatively (e.g., 128 × 128), their applications are therefore limited. In this paper, we introduce a novel SPatial Attention Residual Network (SPARNet) built on our newly proposed Face Units (FAUs) super-resolution. Specifically, spatial attention mechanism vanilla residual blocks. This enables convolutional layers adaptively bootstrap features related key pay less those feature-rich regions. makes training more effective efficient account very small portion image. Visualization maps shows that network capture well even faces 16×16). Quantitative comparisons various kinds metrics (including PSNR, SSIM, identity similarity, detection) demonstrate superiority method over current state-of-the-arts. We further extend SPARNet multi-scale discriminators, named SPARNetHD, produce high results (i.e., 512×512). show SPARNetHD synthetic data not quality outputs synthetically degraded images, but also good generalization ability real world Codes available at https://github.com/chaofengc/Face-SPARNet.