Asymmetric Information Distillation Network for Lightweight Super Resolution

The purpose of this paper is to design a lightweight network to achieve image super resolution performance equivalent to SRResNet. We design an asymmetric information distillation block (AIDB) with distillation information multiplexing and asymmetric information extraction capabilities to better achieve this goal. Distillation information multiplexing refers to the repeated processing of distilled information to supplement the ability of key information extraction. Asymmetric information enhancement block (AIEB) refers to identify different features in the image by the horizontal and vertical feature extraction. AIEB greatly reduces the number of parameters, and distillation information multiplexing works as a supplement to the lost high dimensional information. A large number of experiments show that our asymmetric information distillation network (AIDN) achieves a better balance of performance and complexity than SOTA model. Moreover, Our proposed AIDN ranked second in the model complexity track of NTIRE2022 efficient super resolution challenge. Compared with the first place in this track, we achieves higher PSNR performance on testset with a slight disadvantage in the number of parameters. The code is available at https://github.com/zzksdu/AIDN.

1. Introduction

Single image super resolution (SISR) is a classic low-level computer vision task, which aims to reconstruct a super resolution (SR) image from a single low resolution (LR) image. Since SRCNN [3] was proposed, many deep learning methods [5], [10], [16], [17], [20], [23], [28], [31], [32] have been proposed to improve the performance of image super resolution. SRCNN [3] is the first attempt to reconstruct SR from LR using convolution neural network (CNN). Subsequently, Kim et al. [12] achieved better super resolution performance by increasing the network depth of SRCNN [3], which also shows that the network depth is closely related to the super resolution performance. Bee et al. [16] made greater progress in super resolution by introducing residual structures that further increased the depth of the network.