图像修复问题的低秩对偶逼近算法

Low-rank and dual approximation method for image inpainting problems

为改进传统基于样本修复方法在实际应用中的不足,提出了一种新的图像修复算法.新算法以显著性排序法确保优先修复含明显结构边的目标块,利用图像欧氏距离搜索与该目标块匹配的相似样本块,对由搜索样本向量化构成的相似块矩阵进一步采用低秩对偶逼近提取可用信息以修复缺失像素.实验表明,新算法能够准确地优先修复显著性结构,且对多种类型的缺失均具有较好的修复效果.

A novel image inpainting algorithm is proposed to improve the traditional exemplar-based inpainting methods. The new approach uses a salient-based ranking method to ensure the priority of the target patch with structure edges, whose similar exemplars would be measured by the image Euclidean distance. And a low-rank and dual approximation of the selected exemplar matrix is used to extract the available information for inpainting. Experimental results show that the new algorithm preferentially repairs significant structures accurately, and performs well in different kinds of missing or damaged cases.