Image Up-Sampling for Super Resolution with Generative Adversarial Network

Abstract

In case, we carry out single image Super Resolution (SR) utilizing deep learning, we utilize bicubic interpolation for up-sampling of low resolution images before input them into SR methods. In the preprocessing, these basic interpolation methods cause blur and noise effects for after processed images. These noise images may affect the SR results. In this research, by focusing on this point, we propose a new image up-sampling method utilizing Generative Adversarial Network (GAN). In this work, we improve an image evaluation criterion in generator part of GAN by combining Multi-Scale Structural Similarity (MS-SSIM) and L1 norm. From experiments, we have confirmed that our method allows us to create more qualitatively up-sampling images. As the quantitative results, our proposed method have achieved 0.90 [dB] of average PSNR, 3.35 [%] of average SSIM, and 1.28 [%] of average MS-SSIM improvement using Set 5 and Set 14 dataset compared with bicubic interpolation.