Persistent Memory Residual Network for Single Image Super Resolution

Progresses has been witnessed in single image superresolution in which the low-resolution images are simulated by bicubic downsampling. However, for the complex image degradation in the wild such as downsampling, blurring, noises, and geometric deformation, the existing superresolution methods do not work well. Inspired by a persistent memory network which has been proven to be effective in image restoration, we implement the core idea of human memory on the deep residual convolutional neural network. Two types of memory blocks are designed for the NTIRE2018 challenge. We embed the two types of memory blocks in the framework of enhanced super resolution network (EDSR), which is the NTIRE2017 champion method. The residual blocks of EDSR is replaced by two types of memory blocks. The first type of memory block is a residual module, and one memory block contains four residual modules with four residual blocks followed by a gate unit, which adaptively selects the features needed to store. The second type of memory block is a residual dilated convolutional block, which contains seven dilated convolution layers linked to a gate unit. The two proposed models not only improve the super-resolution performance but also mitigate the image degradation of noises and blurring. Experimental results on the DIV2K dataset demonstrate our models achieve better performance than EDSR.

1. Introduction

Image super-resolution aims at restoring rich details of a high-resolution (HR) image from an LR image or a sequence of low-resolution images without additional hardware support. Moreover, super-resolution (SR) is an ill-posed problem because an LR image can be generated by a large subspace of high-resolution images. Until now, SR is still a challenging task due to the complex degradation in the wild, such as image noises, blurring, downsampling and so on.