计算机工程与应用 ›› 2021, Vol. 57 ›› Issue (2): 230-236.DOI: 10.3778/j.issn.1002-8331.1911-0227

• 图形图像处理 • 上一篇    下一篇

融合邻域回归和稀疏表示的图像超分辨率重构

丁玉祥,卞维新,接标,赵俊   

  1. 1.安徽师范大学 计算机与信息学院,安徽 芜湖 241002
    2.网络与信息安全安徽省重点实验室,安徽 芜湖 241002
  • 出版日期:2021-01-15 发布日期:2021-01-14

Super-Resolution Image Reconstruction Based on Neighborhood Regression and Sparse Representation

DING Yuxiang, BIAN Weixin, JIE Biao, ZHAO Jun   

  1. 1.School of Computer and Information, Anhui Normal University, Wuhu, Anhui 241002, China
    2.Anhui Provincial Key Laboratory of Network and Information Security, Wuhu, Anhui 241002, China
  • Online:2021-01-15 Published:2021-01-14

摘要:

对于图像超分辨率重建而言,通常会将图像的整体信息作为研究对象。然而图像本身含有的大量结构信息并没有得到充分利用。为了提高超分辨率重建的效果,实现对不同特征信息的利用,提出了一种融合邻域回归和稀疏表示的图像超分辨率重构算法。依据图像所具有的低秩性对高分辨率图像进行分解,获得高分辨率图像的低秩部分和稀疏部分;将对应的低分辨率图像与高分辨率图像的低秩部分和稀疏部分进行训练,学习得到对应的特征字典;基于高分辨率图像的低秩部分和稀疏部分分别基于稀疏表示和邻域嵌入进行高分辨率重构;基于低秩矩阵恢复理论,融合邻域回归和稀疏表示重构的高分辨率图像,得到最终的高分辨率图像。在测试集Set5和Set14上将提出的算法与几种经典算法进行对比实验,可视化和量化结果均表明,相比传统超分辨率算法,提出的算法在PSNR和SSIM都有很好的提升。

关键词: 超分辨率, 低秩矩阵恢复, 字典学习, 稀疏表示, 邻域嵌入

Abstract:

For image super-resolution reconstruction, the over-all information of the image is usually the object of study. However, the information of intrinsic structure in the image is not fully utilized. In order to improve the performance of super-resolution reconstruction by using the information of different features, a super-resolution reconstruction algorithm based on neighborhood regression and sparse representation is proposed. Firstly, the high-resolution image is decomposed into the low-rank portion and the sparse portion according to the low rank which the image has. Secondly, the different feature dictionaries are constructed according to the corresponding training samples. The sparse portion and low-rank portion of low-resolution are reconstructed based on the corresponding feature dictionary by using neighborhood regression and sparse representation. Finally, the high-resolution image is obtained based on the reconstructed sparse portion and low-rank portion according to low rank matrix recovery theory. In the test set Set5 and Set14, the proposed algorithm compares with several classical algorithms, and the visualized and quantized results show that compared with the traditional super-resolution algorithm, the proposed algorithm has a good improvement in PSNR and SSIM.

Key words: super resolution, low rank matrix restoration, dictionary learning, sparse representation, neighbor embedding