计算机应用 ›› 2021, Vol. 41 ›› Issue (3): 845-850.DOI: 10.11772/j.issn.1001-9081.2020060979

所属专题: 多媒体计算与计算机仿真

• 多媒体计算与计算机仿真 • 上一篇    下一篇

基于注意力机制的图像超分辨率重建

王永金1, 左羽1,2, 吴恋2, 崔忠伟2, 赵晨洁1   

  1. 1. 贵州大学 计算机科学与技术学院, 贵阳 550025;
    2. 贵州师范学院 数学与大数据学院, 贵阳 550018
  • 收稿日期:2020-07-07 修回日期:2020-10-15 出版日期:2021-03-10 发布日期:2020-12-17
  • 通讯作者: 左羽
  • 作者简介:王永金(1994-),男,湖北荆州人,硕士研究生,主要研究方向:深度学习、超分辨率重构;左羽(1962-),男,贵州贵阳人,教授,硕士,CCF会员,主要研究方向:物联网通信;吴恋(1988-),女,贵州安龙人,副教授,硕士,主要研究方向:深度学习、信息安全;崔忠伟(1980-),男,贵州铜仁人,副教授,博士,主要研究方向:低成本物联网;赵晨洁(1995-),女,河南新乡人,硕士研究生,主要研究方向:病毒检测。
  • 基金资助:
    国家科技部和国家自然科学基金奖励补助基金资助项目(黔科合平台人才[2017]5790-09);贵州省科技计划项目(黔科合基础[2018]1121);贵州省省级重点学科“计算机科学与技术”(ZDXK[2018]007);贵州省教育厅创新群体研究项目(黔教合KY字[2021]022)。

Image super-resolution reconstruction based on attention mechanism

WANG Yongjin1, ZUO Yu1,2, WU Lian2, CUI Zhongwei2, ZHAO Chenjie1   

  1. 1. College of Computer Science and Technology, Guizhou University, Guiyang Guizhou 550025, China;
    2. School of Mathematics and Big Data, Guizhou Education University, Guiyang Guizhou 550018, China
  • Received:2020-07-07 Revised:2020-10-15 Online:2021-03-10 Published:2020-12-17
  • Supported by:
    This work is partially supported by the Award and Subsidy Fund of Ministry of Science and Technology and National Natural Science Foundation of China (QianKeHePingTaiRenCai[2017]5790-09), the Science and Technology Program of Guizhou Province (QianKeHeJiChu[2018]1121), the Guizhou Provincial Key Discipline "Computer Science and Technology"(ZDXK[2018]007), the Guizhou Provincial Department of Education Innovation Group Research Project(QianJiaoHeKYZi[2021]022).

摘要: 目前,单幅图像超分辨率重建取得了很好的效果,然而大多数模型都是通过增加网络层数来达到好的效果,并没有去发掘各通道之间的相关性。针对上述问题,提出了一种基于通道注意力机制(CA)和深度可分离卷积(DSC)的图像超分辨率重建方法。整个模型采用多路径模式的全局和局部残差学习,首先利用浅层特征提取块来提取输入图像的特征;然后,在深层特征提取块中引入通道注意力机制,通过调整各通道的特征图权重来增加通道相关性,从而提取高频特征信息;最后,重建出高分辨率图像。为了减少注意力机制带来的巨大参数影响,在局部残差块中使用了深度可分离卷积技术以大大减少训练参数,同时采用自适应矩估计(Adam)优化器来加速模型的收敛,从而提高了算法性能。该方法在Set5、Set14数据集上进行图像重建,实验结果表明不仅该方法重建的图像具有更高的峰值信噪比(PSNR)和结构相似度(SSIM),而且所提模型的参数量减少为深度残差通道注意力网络(RCAN)模型的参数量的1/26。

关键词: 超分辨率重建, 注意力机制, 深度可分离卷积, 残差网络, 卷积神经网络

Abstract: At present, super-resolution reconstruction of a single image achieves a good effect, but most models achieve the good effect by increasing the number of network layers rather than exploring the correlation between channels. In order to solve this problem, an image super-resolution reconstruction method based on Channel Attention mechanism (CA) and Depthwise Separable Convolution (DSC) was proposed. The multi-path global and local residual learning were adopted by the entire model. Firstly, the shallow feature extraction block was used to extract the features of the input image. Then, the channel attention mechanism was introduced in the deep feature extraction block, and the correlation of the channels was increased by adjusting the weights of the feature graphs of different channels to extract the high-frequency feature information. Finally, a high-resolution image was reconstructed. In order to reduce the huge parameter influence brought by the attention mechanism, the depthwise separable convolution technology was used in the local residual block to greatly reduce the training parameters. Meanwhile, the Adaptive moment estimation (Adam) optimizer was used to accelerate the convergence of the model, so as to improve the algorithm performance. The image reconstruction by the proposed method was carried out on Set5 and Set14 datasets. Experimental results show that the images reconstructed by the proposed method have higher Peak Signal-to-Noise Ratio (PSNR) and Structural SIMilarity index (SSIM), and the parameters of the proposed model are reduced to 1/26 of that of the depth Residual Channel Attention Network (RCAN) model.

Key words: super-resolution reconstruction, attention mechanism, Depthwise Separable Convolution (DSC), residual network, Convolutional Neural Network (CNN)

中图分类号: