基于局部结构相似性的单幅图像超分辨率算法

A New Super-Resolution Algorithm for a Single Image Based on Local Structure Similarity

单幅图像放大是一个病态问题,约束信息的不足令问题的解决难以有突破性的进展基于图像局部结构的相似性和在不同分辨率尺度上的相似保持,提出了利用局部结构相似性来约束超分辨率问题的思路 根据相似判断规则发现图像中的相似区域,根据相似程度生成多幅相似图像,从而可以利用图像序列超分辨率的算法来求解 文中使用最大后验概率估计方法,在最大后验概率意义下得到最优解实验表明,该算法用于存在着大量相似结构的图像超分辨率问题中有着很好的效果,譬如文字图像。

Single-image zooming is an ill-posed problem due to constrained information which may be provided. This paper pays attention to the similarity feature among local structures in an image which can be maintained across scale. Based on the feature, we propose a new method using similarity. By the method, first, we try to find out all the similar structure sets under certain similarity criterion and obtain the degree of similarity. Then according to the degree sorted, image sequences in similarity are generated. As a result, we can apply known algorithms in the field of image sequence super-solution to solve the problem. This paper selects the MAP method and computes the optimal resolution by the steep-descending iterations. Several experiments are presented to demonstrate the effectiveness of the approach, especially in the area of IC, where the images are often with plenty of similar structures.