Encoder-Decoder Residual Network for Real Super-Resolution

Real single image super-resolution is a challenging task to restore lost information and attenuate noise from images mixed unknown degradations complicatedly. Classic single image super-resolution, aims to enhance the resolution of bicubically degraded images, has recently obtained great success via deep learning. However, these existing methods do not perform well for real single image super-resolution. In this paper, we propose an Encoder-Decoder Residual Network (EDRN) for real single image super-resolution. We adopt an encoder-decoder structure to encode highly effective features and embed the coarse-to-fine method. The coarse-to-fine structure can gradually restore lost information and reduce noise effects. We empirically rethink and discuss the usage of batch normalization. Compared with state-of-the-art methods in classic single image super-resolution, our EDRN can efficiently restore the corresponding high-resolution image from a degraded input image. Our EDRN achieved the 9th place for PSNR and top 5 for SSIM in the final result of NTIRE 2019 Real Super-resolution Challenge. The source code and the trained model are available at https://github.com/yyknight/NTIRE2019_EDRN.

1. Introduction

Single image super-resolution (SISR) is a fundamental low-level vision task in computer vision. The purpose of SISR is to reconstruct the corresponding high-resolution image from a low-resolution image given. Super-resolution technologies have been leveraged in a wide range of fields such as remote sensing [23], satellite imaging [33], and medical imaging [30]. However, since there are plenty of solutions for any single input image, SISR is a highly ill-posed inverse problem. To model the inverse mapping, numerous methods for SISR, including deep-learning based ones, have been proposed. Deep learning based SISR methods have developed explosively in recent years [5], [6], [14], [15], [41]. These methods are effective for the bicubic degradation. However, the degradations for real-world low-resolution images are blind. Therefore, the recent methods cannot accurately restore real-world low-resolution images. Figure 1.

Our result on “cam2_09”, an image comes from NTIRE 2019 real super-resolution challenge. Our result can restore abundant high-frequency details. + denotes the result with self-ensemble.