Face Super-Resolution Using Recurrent Generative Adversarial Network

Abstract

Theface super-resolution (SR) networks based on deep learning have more advanced performance than traditional SR algorithms. However, facial key components are difficult to reconstruct because the adjacent pixels have great change. Moreover, most face SR networks only focus on the performance and ignore the number of parameters. To solve the above problems, we propose the face super-resolution network using recurrent generative adversarial network (FSRRGAN). The generator is the face SR recurrent generator (FSRRG) with dense iterative up-down sampling blocks as the basic unit. It can reduce the number of parameters and effectively improve the reconstruction performance combined with the relativistic average patch discriminator (RAPD). We use the facial perceptual similarity distance (FPSD) loss to replace the traditional perceptual loss. The experimental results show that our network has excellent performance both qualitatively and quantitatively on 4x and 8x face reconstruction.