《计算机应用研究》|Application Research of Computers

改进的单幅图像自学习超分辨率重建方法

Improved super-resolution reconstruction method for self-learning of single image

免费全文下载 (已被下载 次)  
获取PDF全文
作者 王晓明,黄凤,刘少鹏,徐涛
机构 西华大学 a.计算机与软件工程学院;b.机器人研究中心,成都 610039
统计 摘要被查看 次,已被下载
文章编号 1001-3695(2019)08-062-2534-05
DOI 10.19734/j.issn.1001-3695.2018.02.0184
摘要 针对传统超分辨率重建方法稀疏表示依赖大训练样本字典的局限性问题,基于<i>L</i><sub>2</sub>范数的弱稀疏性特点,提出一种改进的单幅图像自学习超分辨率重建方法。通过自学习建立非金字塔阶梯式训练图像集,采用自定义的方法分别提取训练集中低分辨率和相应高分辨率图像特征块及特征像素值;结合<i>L</i><sub>2</sub>范数的协作表示(collaborative representation,CR)理论和支持向量回归(support vector regression,SVR)技术学习多层超分辨率映射模型。实验结果表明,提出的超分辨率方法不仅可行有效,而且与传统的单幅图像的超分辨率方法比较,其PSNR平均提高了0.06~3.92 dB,SSIM平均提高了0.002 4~0.034 8,从客观数值和主观视觉证明了所提方法的优秀性。
关键词 单幅图像超分辨率; <;i>;L<;/i>;<;sub>;2<;/sub>;范数; 协作表示; 支持向量回归
基金项目 国家教育部春晖计划资助项目(Z2015102)
国家自然科学基金资助项目(61532009)
四川省教育厅自然科学重点项目(11ZA004)
本文URL http://www.arocmag.com/article/01-2019-08-062.html
英文标题 Improved super-resolution reconstruction method for self-learning of single image
作者英文名 Wang Xiaoming, Huang Feng, Liu Shaopeng, Xu Tao
机构英文名 a.School of Computer & Software Engineering,b.Robotics Research Center,Xihua University,Chengdu 610039,China
英文摘要 Aiming at the limitation of sparse representation depended on large training sample dictionaries for traditional super-resolution reconstruction method, this paper proposed an improved super-resolution reconstruction method for self-learning of single image based on the characteristic of <i>L</i><sub>2</sub>-norm's weak sparsity. Firstly, it used self-learning to establish the non pyramid stepped training images. Then, it used the custom method to extract feature blocks and feature pixel values of corresponding LR and HR images. Finally, combined with the CR theory of <i>L</i><sub>2</sub>-norm and SVR technology, it established mapping model of the super-resolution. Experimental results show that the proposed super-resolution method is feasible and effective. The average PSNR increases for 0.06~3.92 dB and SSIM increases for 0.002 4~0.034 8 compared with other conventional super-resolution approaches of single image. From the objective and subjective vision, it is proved that the proposed method is excellent.
英文关键词 super-resolution of single image; < i> L< /i> < sub> 2< /sub> -norm; collaborative representation; support vector regression
参考文献 查看稿件参考文献
 
收稿日期 2018/2/3
修回日期 2018/3/29
页码 2534-2538
中图分类号 TP391.41
文献标志码 A