基于聚类分割和纹理合成的图像修复改进算法

Improved algorithm for image inpainting based on clustering segmentation and texture synthesis

Criminisi提出的基于样本的图像修复技术需要在整幅图像中遍历样本,代价太大,并可能因选择错误的样本,不断迭代更新后而导致错误信息累积,使修复结果出现较大的偏差。同时,考虑到Criminisi算法中优先权函数的计算失误可能导致修复结果中出现结构失真,由此提出一种基于聚类分割和纹理合成的图像修复改进算法,将目标样本块的搜索限定在与源样本块所覆盖的类别一致的区域当中。在像素点优先权计算中,引入该像素点邻域灰度梯度差值信息,提出更为合理的优先权计算公式,以最大限度保证复杂场景中边缘优先传递,并在置信度更新项中有差别地对待新填充像素点。通过实验证明,改进算法不仅解决了Criminisi算法可能存在的结构偏差延续问题,修复视觉效果更加符合人们的主观感受,而且大大缩短了修复时间。

Criminisi proposed exemplar-based image inpainting techniques need to traverse the whole image exemplar, it is too costly, and may choose the wrong exemplar, constantly updates iteration error messages resulting cumulative, so that a greater deviation may be in inpainting results. Meanwhile, considering the Criminisi algorithm priority function cal-culation may lead to a structural distortion in inpainting results, which proposes an improved algorithm for image inpainting based on clustering segmentation and texture synthesis, the search will be limited to the same categories zone with the source exemplar covered. In the pixel priority calculation, the pixel neighborhood gray gradient difference information is introduced, the priority of more reasonable formula is proposed to ensure maximum edge preferentially transmitted in complex scenes and update entries in confidence difference to treat newly filled pixels. The experimental results show that the improved algorithm not only solves the Criminisi algorithm possible continuation of structural bias problem, repairing the visual effect is more in line with people’s subjective feelings, but also greatly shortens the repair time.