计算机应用 ›› 2018, Vol. 38 ›› Issue (3): 854-858.DOI: 10.11772/j.issn.1001-9081.2017081920

• 虚拟现实与多媒体计算 • 上一篇    下一篇

结合结构自相似性和卷积网络的单幅图像超分辨率

向文, 张灵, 陈云华, 姬秋敏   

  1. 广东工业大学 计算机学院, 广州 511400
  • 收稿日期:2017-08-07 修回日期:2017-10-12 出版日期:2018-03-10 发布日期:2018-03-07
  • 通讯作者: 向文
  • 作者简介:向文(1993-),男,湖南怀化人,硕士研究生,主要研究方向:图像处理、模式识别;张灵(1968-),女,广西南宁人,教授,博士,主要研究方向:数据挖掘、计算机视觉、无线传感;陈云华(1977-),女,湖北仙桃人,讲师,博士,主要研究方向:计算机视觉、模式识别、深度学习;姬秋敏(1992-),女,山东济宁人,硕士研究生,主要研究方向:图像处理、模式识别。
  • 基金资助:
    广东省自然科学基金博士启动项目(2014A030310169);广东省自然科学基金面上项目(2016A030313703,2016A030313713);广东省科技计划项目(2016B030305002);广东省交通运输厅科技项目(科技-2016-02-030)。

Single image super resolution combining with structural self-similarity and convolution networks

XIANG Wen, ZHANG Ling, CHEN Yunhua, JI Qiumin   

  1. School of Computers, Guangdong University of Technology, Guangzhou Guangdong 511400, China
  • Received:2017-08-07 Revised:2017-10-12 Online:2018-03-10 Published:2018-03-07
  • Supported by:
    This work is partially supported by the Guangdong Natural Science Foundation Doctoral Program (2014A030310169), the Natural Science Foundation Project of Guangdong Province (2016A030313703, 2016A030313713), the Science and Technology Project of Guangdong Province (2016B030305002), the Science and Technology Project of Guangdong Department of Transportation (Sci. and Tech.-2016-02-030).

摘要: 针对单幅图像超分辨率(SR)复原病态逆问题,在重建过程边缘细节丢失导致的模糊,提出一种结合结构自相似和卷积网络的单幅图像超分辨率算法。首先,通过将尺度分解获得待重构图片样本的自身结构相似性,结合外部数据库样本结合作为训练样本,可以解决样本过于分散的问题;其次,将样本输入卷积神经网络(CNN)进行训练学习,得到单幅图像超分辨率的先验知识;然后,利用非局部约束项自适应选择最优字典重建图像;最后,利用迭代反投影算法对图像超分辨率效果进一步提升。实验结果表明,与双三次插值(Bicubic)方法、K-SVD算法和基于卷积神经网络的图像超分辨率(SRCNN)方法等优秀算法相比,所提算法可以得到边缘更为清晰的超分辨率重建效果。

关键词: 超分辨率, 结构自相似性, 深度卷积网络, 正则化, 块匹配

Abstract: Aiming at the ill-posed inverse problem of single-image Super Resolution (SR) restoration, a single image super resolution algorithm combining with structural self-similarity and convolution networks was proposed. Firstly, the self-structure similarity of samples to be reconstructed was obtained by scaling decomposition. Combined with external database samples as training samples, the problem of over-dispersion of samples could be solved. Secondly, the sample was input into a Convolution Neural Network (CNN) for training and learning, and the prior knowledge of the super resolution of the single image was obtained. Then, the optimal dictionary was used to reconstruct the image by using a nonlocal constraint. Finally, an iterative backprojection algorithm was used to further improve the image super resolution effect. The experimental results show that compared with the excellent algorithms such as Bicubic, K-SVD (Singular Value Decomposition of k iterations) algorithm and Super-Resolution Convolution Neural Network (SRCNN) algorithm, the proposed algorithm can get super-resolution reconstruction with clearer edges.

Key words: Super Resolution (SR), structural self-similarity, deep convolution network, regularization, block matching

中图分类号: