SIEN:Spatial Information Enhancement Network for Blind Super-Resolution

Blind super-resolution methods based on deep learning have attracted increasing interest due to their excellent performance. However, most existing blind SR models underutilize the contextual information of features in LR images, thus relatively reducing the texture details of the images. In this paper, we propose a spatial information enhancement network (SIEN) to address this problem in image blind SR. The network adds a new branch SET module to the idea of alternate iterative optimization of image and blur kernels in blind SR, which acquires global contextual information by making full use of LR image features and restores texture details based on the contextual information. And after extensive experiments, it is proved that our proposed method can have more accurate SR results under the real world.

1 Introduction

As a fundamental problem in computer vision [1], super-resolution has been a popular research topic and has direct or indirect applications in many fields. After years of development, the overall approach of super-resolution methods can usually be divided into three categories namely interpolation-based methods, model-based methods, and deep learning-based methods, where deep learning-based methods tend to have better results compared to the first two, and deep learning-based methods usually have better results in terms of feature representation as well as stronger inference capabilities and easier to implement end-to-end training frameworks. As a result, deep learning has been dominant in solving super-resolution problems for the past many years. However, most of these methods rely heavily on fixed and known downsampling methods [6], [8], [9], [10], and this limitation may hinder the performance of these methods when applied to realistic scenarios with unknown or uncertain degradation patterns. To address this challenge, researchers have proposed blind super-resolution (SR) methods designed to handle unknown or uncertain degenerate patterns.