Enhanced Discriminative Generative Adversarial Network for Face Super-Resolution

摘要

Recently, some generative adversarial network (GAN)-based super-resolution (SR) methods have progressed to the point where they can produce photo-realistic natural images by using a generator (G) and discriminator (D) adversarial scheme. However, vanilla GAN-based SR methods cannot achieve good reconstruction and perceptual fidelity on real-world facial images at the same time. Because of D loss, them are hard to converge stably, which may cause the model collapse. In this paper, we present an Enhanced Discriminative Generative Adversarial Network (EDGAN) for SR facial recognition to achieve better reconstruction and perceptual fidelities. First, we discover that a versatile D boosts the adversarial framework to a preferable Nash equilibrium. Then, we design the D via dense connections, which brings more stable adversarial loss. Furthermore, a novel perceptual loss function, by reusing the intermediate features of D, is used to eliminate the gradient vanishing problem of Gs. To our knowledge, this is the first framework to focus on improving the performance of the D. Quantitatively, experimental results show the advantages of EDGAN on two widely used facial image databases against the state-of-the-art methods with different terms. EDGAN performs sharper and realistic results on real-world facial images with large pose and illumination variations than its competitors.