Abstract
Recent progress in the field of high-fidelity image synthesis using the generative adversarial networks is very promising such as BigGAN. However, extending BigGAN towards image super-resolution still poses challenges: (1) how to retain the fine texture as of original high-resolution images; (2) develop a lightweight model to allow giant GAN model to work in an power efficient manner and friendly on edge devices. In this work, we develop a SR-BigGAN with priors and knowledge distillation (KD) for efficient image super-resolution (SR). First, our new model is an extension of BigGAN to deep SR pipeline, retaining both generator and discriminator architecture with modifications to accommodate SR pipeline. Second, prior knowledge from the low-resolution images are fully leveraged to refine the SR process. Third, to reduce the computational cost, we are distilling the knowledge of trained model, namely, teacher network (TN) into student network (SN) which is more efficient and compact. Extensive experiments on SR datasets, including DIV2K, Imagenet-Mini, Set5, Set14 and Pascal were evaluated in an attempt to compare the Structural Similarity (SSIM), Peak Signal-to-Noise Ratio (PSNR) with the state-of-the-art models to show that there exists significant perceptual enhancement of synthesized images from our new model that directly influences classifiers with enhanced classification.