Single image super-resolution with enhanced Laplacian pyramid network via conditional generative adversarial learning

Despite much progress has been made by applying generative adversarial network (GAN) to single image super-resolution (SISR), obvious difference remains between the details of reconstructed high-frequency and ground-truth because GAN is unstable that has a very high degree of freedom. To address this issue, we exploit conditional GAN (CGAN) for SISR, which leverages the ground-truth high-resolution (HR) image as its conditional variable to guide to learn a more stable model. To better reconstruct image with a large-scale factor, we further design an enhanced Laplacian pyramid network (ELapN) as the generator model of CGAN, which progressively reconstructs HR images at multiple pyramid levels. The proposed ELapN fuses low-and high-level features for the residual image learning achieves better generalization than those only based on high-level information. Finally, we train the proposed network via deep supervision using a combination of multi-level CGAN, VGG and robust Charbonnier loss functions to obtain high-quality SR results. Extensive evaluations on three benchmark datasets including Set5, Set14, B100 demonstrate superiority of the proposed method over state-of-the-art methods in terms of PSNR, SSIM and visual effect.