电子学报 ›› 2022, Vol. 50 ›› Issue (10): 2336-2346.DOI: 10.12263/DZXB.20201089

• 学术论文 • 上一篇    下一篇

一种轻量级的多尺度通道注意图像超分辨率重建网络

周登文, 李文斌, 李金新, 黄志勇   

  1. 华北电力大学控制与计算机工程学院,北京 102206
  • 收稿日期:2020-09-30 修回日期:2021-05-14 出版日期:2022-10-25
    • 作者简介:
    • 周登文 男,1965年出生,湖北人.华北电力大学控制与计算机工程学院教授.主要研究方向为神经网络和深度学习在图像处理和计算机视觉中的应用,特别是图像超分辨率技术.E-mail: zdw@ncepu.edu.cn
      李文斌 男,1996年出生,甘肃人.2017年在上海电力学院获得学士学位,现为华北电力大学硕士研究生.主要研究方向为计算机视觉和深度学习.E-mail: 1182227108@ncepu.edu.cn

Image Super-Resolution Reconstruction Based on Lightweight Multi-Scale Channel Attention Network

ZHOU Deng-wen, LI Wen-bin, LI Jin-xin, HUANG Zhi-yong   

  1. School of Control and Computer Engineering,North China Electric Power University,Beijing 102206,China
  • Received:2020-09-30 Revised:2021-05-14 Online:2022-10-25 Published:2022-10-11

摘要:

近年来,基于深度卷积神经网络的图像超分辨率技术取得了突出进展,并主导了当前的超分辨率技术的研究.但是,性能的改进,往往以参数量的急剧增加为代价,这限制了超分辨率方法的实际应用.本文设计了一个轻量级单图像超分辨率深度卷积网络,主要贡献包括:提出了一个多尺度的特征融合模块,使用不同感受野的卷积核,提取多种尺度的特征;提出了一个通道搅乱注意力模块,促进特征通道之间的信息流动,并增强特征选择能力;提出了一个全局特征融合连接模块,提高浅层特征的利用率.实验证明,本文方法与当前代表性的方法MSRN(Multi-Scale Residual Network)相比,参数量减少了3/4,重建的高分辨率图像的主观和客观质量均显著更好.

关键词: 超分辨率, 深度学习, 卷积神经网络, 注意力机制, 多尺度特征

Abstract:

Recently, image super-resolution technology based on deep convolutional neural network has made remarkable achievements and has become popular in the current super-resolution technology. However, superior performance is often at the expense of the large number of parameter amounts, which limits the real-world applications for single image super-resolution. In this paper, a lightweight single image super-resolution deep convolutional network is proposed. The main contributions of this paper are as follows: a multi-scale feature fusion block is proposed to extract multiple features via convolution kernels with different receptive fields; the channel shuffle attention mechanism we designed promotes the flow of the information across feature channels, which enhances the ability of feature selection; a global feature fusion connection is proposed to improve the feature utilization. Extensive experiments demonstrate that the parameter amounts of our method reduced by 3/4 compared with the current state-of-the-art MSRN method, while subjective visual and objective quality of the reconstructed high-resolution image are perform significantly better.

Key words: super-resolution, deep learning, convolutional neural network, attention mechanism, multi-scale feature

中图分类号: