Frequency Separation Network for Image Super-Resolution

Shanshan Li,Haisheng Li,Jian Cao,Qiang Cai,Zhuangzi Li,Lei Wang

doi:10.1109/access.2020.2972927

Abstract

It is well-known that high-frequency information (e.g. textures, edges) is significant for single image super-resolution (SISR). However, Existing of deep Convolutional Neural Network (CNN) based methods directly model mapping function from low resolution (LR) to high resolution (HR), and they treat high-frequency and low-frequency information equally during feature extraction. Therefore, the high-frequency learning mode can not be sufficiently attentive, resulting in inaccurate representation of some local details. In this study, we aim to build potential frequencies' relations and handle high-frequency and low-frequency information differentially. Specifically, we propose a novel Frequency Separation Network (FSN) for image super-resolution (SR). In FSN, a new Octave Convolution (OC) is adopted, which uses four operations to perform information update and frequency communication between high frequency and low frequency features. In addition, global and hierarchical feature fusion are employeed to learn elaborate and comprehensive feature representations, in order to further benefit the quality of final image reconstruction. Extensive experiments conducted on benchmark datasets demonstrate the state-of-the-art performance of our method.

Highlights

single image super-resolution (SISR) aims at producing a visually high resolution (HR) output from its low resolution (LR) input, and it has recently attracted much attention in academic and industrial communities
To practically tackle the above-mentioned problem, we propose a novel image SR framework based on Frequency Separation Network (FSN)
We propose a novel FSN for image super-resolution, which respectively embed LR image into high-frequency and low-frequency feature space for more accurate image reconstruction

Summary

Introduction

SISR aims at producing a visually HR output from its LR input, and it has recently attracted much attention in academic and industrial communities. SISR is an inherently ill-posed problem since the mapping from the LR to HR space has multiple solutions. For solving this issue, various promising SR methods [2]–[9] have been proposed in the past years. A series of methods [12]–[16] construct deeper architectures to obtain more informative and multifarious feature representations, in order to benefit high-resolution reconstruction. These methods do not consider the latent frequency relations of input LR image. The low-frequency information can describe the overall outline of one image

Objectives

Methods

Conclusion