计算机工程

• 图形图像处理 • 上一篇    下一篇

频域内基于邻域特征学习的单幅图像超分辨重建

苏富林  1,钱素娟  2,魏霖静  3,孙连海  4   

  1. (1.甘肃民族师范学院 计算机科学系,甘肃 合作 747000; 2.郑州财经学院 信息工程学院,郑州 450000; 3.甘肃农业大学 信息科学技术学院,兰州 730070; 4.成都师范学院 计算机科学学院,成都 611130)
  • 收稿日期:2016-03-22 出版日期:2017-05-15 发布日期:2017-05-15
  • 作者简介:苏富林(1974—),男,副教授、硕士,主研方向为图像处理、数据挖掘、神经网络;钱素娟(通信作者),副教授、硕士;魏霖静,副教授、博士后;孙连海,实验师、硕士。
  • 基金项目:

    国家自然科学基金(034031122, 61063028);甘肃省教育厅项目(2014A-115)。

Single-image Super-resolution Reconstruction in Frequency Domain Based on Neighborhood Feature Learning

SU Fulin  1,QIAN Sujuan  2,WEI Linjing  3,SUN Lianhai  4   

  1. (1.Department of Computer Science,Gansu Nomal University for Nationalities,Hezuo,Gansu 747000,China; 2.College of Information Engineering,Zhengzhou Institute of Finance Economics,Zhengzhou 450000,China; 3.School of Information Science and Technology,Gansu Agricultural University,Lanzhou 730070,China; 4.School of Computer Science,Chengdu Normal University,Chengdu 611130,China)
  • Received:2016-03-22 Online:2017-05-15 Published:2017-05-15

摘要:

针对图像重建过程中待插值点灰度估计不准确的问题,提出一种基于邻域特征学习的单幅图像超分辨回归分析方法。在输入低分辨率图像后,利用图像特征从低分辨率图像及其对应高分辨率图像的几何相似结构中学习局部协方差。对于邻域中的每一个图像块,估计4个方向的方差以适应插值像素。实验结果表明,该方法既能保证重建的高分辨率图像均匀区域的一致性,同时也能完整保留图像细节信息和边缘轮廓。

关键词: 协方差矩阵, 方向方差, 傅里叶特征, 插值, 核回归, 超分辨重建

Abstract:

Aiming at the problem that the interpolation point gray estimation is not accurate in the process of image reconstruction,a single-image super-resolution algorithm in frequency domain based on neighborhood feature learning is proposed in this paper.When giving a low-resolution image as input,it uses image feature to learn local covariance from low-resolution image and its corresponding high-resolution image’s geometric similar structure.For each patch in the neighborhood,four directional variances are estimated to adapt the interpolated pixels.Experimental results demonstrate that the proposed method not only can guarantee the consistency of the smooth region in the reconstructed high-resolution image,but also can retain the image details and the integrity of the edge profile.

Key words: covariance matrix, direction variance, Fourier feature, interpolation, kernel regression, super-resolution reconstruction

中图分类号: