Abstract: In order to solve the problem of inaccurate filling orders in Criminisi algorithm, an exemplar-based algorithm for image inpainting is proposed. Color variation between neighbor pixels is used to define the priority function and image structures propagate firstly. Besides, quality of exemplar filling is introduced into the update of confidence term. Confidence value is relatively high when the difference between the target patch and the similar patch is small. In addition, the patch that with minimum gradient distance is adopted when there are several most similar blocks with a same color distance. Experimental results show that the improved algorithm achieves high inpainting quality on various kinds of images.

Key words: image inpainting, texture synthesis, priority, gradient distance, update of confidence term

摘要： 在Criminisi算法基础上，针对其填充次序不准确等缺点，提出了一种基于样例的图像修复方法。通过在优先权函数中引入相邻像素间颜色差值信息，图像结构部分优先传播。同时，置信度值更新考虑样本填充质量，目标块与相似块差异越小则该值相对较高。另外，块搜索过程中多个样本块有相同最小颜色距离时，选取梯度距离最小的样本块。实验结果表明，改进算法对各类图像取得了较高修复质量。

关键词: 图像修复, 纹理合成, 优先权, 梯度距离, 置信度项更新