基于双字典正则化的单帧图像超分辨率重建方法  

Learning dual dictionary regularization for single image super-resolution

作　　者：崔琛 张凯兵[1] CUI Chen;ZHANG Kaibing(School of Electronics and Information, Xi’an Polytechnic University, Xi’an 710048, China)

机构地区：[1]西安工程大学电子信息学院,陕西西安710048

出　　处：《西安工程大学学报》2021年第2期66-72,共7页Journal of Xi’an Polytechnic University

基　　金：国家自然科学基金(61971339,61471161);陕西省自然科学基础研究计划重点项目(2018JZ6002);西安工程大学博士启动基金(BS1616)。

摘　　要：为提高单帧图像的分辨率,提出一种基于内部字典和外部字典正则化的超分辨(super-resolution,SR)重建方法。首先,将输入的低分辨率(low-resolution, LR)图像划分为若干个结构相似的子区域,对每个子区域采用主成分分析(principal components analysis, PCA)字典学习方法构造每个子区域对应的内部字典;其次,将外部高分辨率(high-resolution, HR)图像的高频细节分为结构相似组,采用PCA字典学习方法构造外部字典;再次,使用非局部回归模型设计2个具有互补性的正则化项用于解决SR不确定性问题;最后,采用梯度下降迭代优化算法实现SR重构。实验结果表明,相比于对比方法,算法的峰值信噪比(peak signal to noise ratio, PSNR)平均提升0.2 dB,结构相似度(structural similarity, SSIM)平均提升0.01,并且能够获得更好的主观视觉效果。In this paper, a super-resolution (SR) method through an internal and external dictionary-based regularization was proposed for improving the resolution of single low-resolution (LR) image.Firstly, the input was divided into several similar groups, and then the principal components analysis (PCA) dictionary learning method was used to construct the internal dictionary of each corresponding group.Secondly, the high-frequency details of external high-resolution (HR) images was divided into groups with similar structures, and the PCA method was adopted to construct the external dictionary corresponding to each group.Next, the nonlocal regression model was used to design two complementary regularities to solve the problem of SR uncertainty.Finally, the SR reconstruction was achieved through an iterative optimization algorithm with gradient descent.The experimental results showed that the proposed method could improve the peak signal to noise ratio (PSNR) by 0.2 dB on average and the structural similarity (SSIM) by 0.01 on average, showing better visual quality than the other compared methods.

关 键 词：图像超分辨重建 正则化 可控核回归 局部字典学习 非局部相似性 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]