基于多尺度非局部约束的单幅图像超分辨率算法

多尺度结构自相似性是指图像中的大量物体具有相同尺度以及不同尺度相似结构的性质.本文提出了一种基于多尺度非局部约束的单幅图像超分辨率算法,结合多尺度非局部方法和多尺度字典学习方法将蕴含在图像多尺度自相似结构中的附加信息加入到重建图像中.多尺度非局部方法在图像金字塔的不同层中搜索相似图像块,并利用多尺度相似图像块间的关系建立非局部约束项,通过正则化约束获取多尺度自相似结构中的附加信息;多尺度字典学习方法将图像金字塔作为字典学习的样本,通过字典学习使样本中的多尺度相似图像块 在字典下具有稀疏表示形式,从而获取多尺度自相似结构中的附加信息.实验表明, 与ScSR、SISR、NLIBP、CSSS、ASDSAR和mSSIM等算法相比,本文的算法取得了更好的超分辨率重建效果.

Single-image Super-resolution Algorithm Based on Multi-scale Nonlocal Regularization

Multi-scale structural self-similarity refers to that there are many similar structures in the same image, which are either in the same scale or across different scales. In this paper, a single-image super-resolution method based on multi-scale nonlocal regularization is proposed. In this method, the multi-scale nonlocal and the multi-scale dictionary learning methods are combined to add the extra information exploited from multi-scale similar structures into the reconstructed image. The multi-scale nonlocal method exploits extra information from multi-scale similar structures by searching for similar patches in the image pyramid and constructing the multi-scale nonlocal regularization according to the correspondence between multi-scale similar patches. The multi-scale dictionary learning method exploits extra information from multi-scale similar structures by using the image pyramid as training samples in dictionary learning, so that the patches in the pyramid have sparse representations over the learned dictionary. Experimental results demonstrate that the method achieves better image quality compared with ScSR, SISR, NLIBP, CSSS, ASDSAR and mSSIM methods.