J4 ›› 2014, Vol. 41 ›› Issue (4): 100-107.doi: 10.3969/j.issn.1001-2400.2014.04.018

• 研究论文 • 上一篇    下一篇

广义非局部均值和自相似性的超分辨率算法

吴炜1;郑成林2;张莹莹1;周寿桓1   

  1. (1. 四川大学 电子信息学院,四川 成都  610064;
    2. 华为技术有限公司,广东 深圳  518129)
  • 收稿日期:2013-04-11 出版日期:2014-08-20 发布日期:2014-09-25
  • 通讯作者: 吴炜
  • 作者简介:吴炜(1975-), 男, 副教授, 博士, E-mail:wuwei@scu.edu.cn.
  • 基金资助:

    国家自然科学基金资助项目(61271330);中国科学院数字地球重点实验室资助项目(2012LDE016);武汉大学测绘遥感信息工程国家重点实验室资助项目(12R03);华为技术有限公司资助项目;四川省重大科技支撑计划资助项目(2014GZ0005);博士后基金资助项目(2014M552357)

Image super-resolution using generalized nonlocal mean and self-similarity

WU Wei1;ZHENG Chenglin2;ZHANG Yingying1;ZHOU Shouhuan1   

  1. (1. College of Electronics and Information Engineering, Sichuan Univ., Chengdu  610064,China;
    2. Huawei Technologies Co., Ltd., Shenzhen  518129,China)
  • Received:2013-04-11 Online:2014-08-20 Published:2014-09-25
  • Contact: WU Wei

摘要:

提出一种利用广义非局部均值和自相似性的图像超分辨率算法.该算法不仅利用图像的自相似性将低分辨率图像与其下采样图像作为一个训练库,而且利用非局部平均算法的良好特性提高复原图像的质量.该算法首先提取低分辨率图像的高斯差特征系数,然后利用广义非局部平均算法来估计待复原图像丢失的高频细节,获得高分辨率图像.实验结果表明,该算法对图像取得较好的复原效果,复原出的高分辨率图像更接近于真实图像,与其他方法相比,具有更好的主观和客观质量.

关键词: 图像复原, 图像处理, 基于学习的超分辨率, 非局部平均算法

Abstract:

A super-resolution method based on generalized nonlocal mean and self-similarity is proposed. The proposed method not only adopts the self-similarity of the image by taking the low-resolution image and its downsampled version as a training set but uses the nonlocal mean algorithm to improve the quality of the restored image. The proposed method first extracts the features of the low image by using the difference of Gaussians, and then a generalized nonlocal mean algorithm is adopted to estimate the high-frequency details of the low image. Experimental results show that the proposed algorithm has a good performance, and that the high-resolution image generated by the proposed method is of better subjective and objective quality compared with other methods.

Key words: image restoration, image processing, learning-based super-resolution, nonlocal means

中图分类号: 

  • TP391.4