基于深度学习特征字典的单帧图像超分辨率重建
作者:
作者单位:

作者简介:

赵丽玲(1978-),女,博士研究生,研究方向:图像处理与超分辨技术,E-mail:zhaoliling@nuist.edu.cn;孙权森(1963-),男,教授,研究方向:模式识别、图像处理、遥感信息系统;张泽林(1993-),男,硕士研究生,研究方向:深度学习理论及应用

通讯作者:

基金项目:

国家自然科学基金(61673220,61802199)资助项目。


Single Image Super Resolution Reconstruction Based on Deep Features Dictionary
Author:
Affiliation:

Fund Project:

  • 摘要
  • |
  • 图/表
  • |
  • 访问统计
  • |
  • 参考文献
  • |
  • 相似文献
  • |
  • 引证文献
  • |
  • 资源附件
    摘要:

    在基于字典的单帧图像超分辨率重建算法中,依赖人工浅层特征设计的字典表达图像特征能力有限。为此,提出基于深度学习特征字典的超分辨重建方法。该算法首先利用深度网络进行高、低分辨率训练样本图像深层次特征学习;然后,在稀疏字典超分辨框架下联合训练特征字典;最后,输入单帧低分辨率图像并利用该字典实现超分辨率重建。理论分析表明,引入深度网络提取图像深层次特征并用于字典训练,对低分辨率图像的高频信息补充更加有利。实验证明,与双三次插值以及基于一般人工特征字典的超分辨重建算法相比,本文算法的主观视觉和客观评价指标均高于对比算法。

    Abstract:

    The ability of image features expression with the dictionary designed by artificial shallow features is limited in dictionary based single image super-resolution reconstruction algorithm. For the reason, an image super resolution reconstruction algorithm based on deep learning and feature dictionary is proposed. Firstly, deep-level feature learning is carried out in high and low resolution training sample images by using deep network. Secondly, the feature dictionary is trained with the combination of sparse coding under the sparse dictionary super resolution frame. Finally, a low resolution image is put in and the super resolution reconstruction is realized by using the dictionary. Theoretical analysis shows that the combination of image deep-level feature extraction and dictionary training by using deep network is more beneficial to high frequency information supplement for low resolution image. Experimental results show that compared with bicubic interpolation and other general artificial feature dictionary based super resolution reconstruction algorithms, the proposed algorithm has better subjective visual and objective evaluation indices.

    参考文献
    相似文献
    引证文献
引用本文

赵丽玲, 孙权森, 张泽林.基于深度学习特征字典的单帧图像超分辨率重建[J].数据采集与处理,2018,33(4):740-750

复制
分享
文章指标
  • 点击次数:
  • 下载次数:
历史
  • 收稿日期:2016-10-03
  • 最后修改日期:2017-10-09
  • 录用日期:
  • 在线发布日期: 2018-09-08