利用大位移视图修复单幅图像的透视畸变优化算法

Optimization of Perspective Distortion for Image Completion Based on the Large Displacement View

本文提出一个利用大位移视图可靠地恢复目标图像上含有复杂结构信息的较大信息丢失区域的图像修复新算法.依次经过基于单应矩阵的大位移视图变形、基于能量优化的重叠像素对应和基于能量优化的丢失像素估计三个步骤,一个由粗到细的畸变校正算法对大位移视点图像进行透视畸变校正后再用来补全目标图像上的受损区域.最后,我们采用泊松图像融合算法消除修复区域与其周围像素之间存在的鬼影现象,获得无缝的修复效果.实验证明本文方法优于已有的图像修复算法.

We present a new image completion method based on a large displacement view（ LDV） for faithfully repairing large missing regions on the target image with complex structure information. A coarse-to-fine distortion correction algorithm is proposed to minimize the perspective distortion in the LDV image, which performs following three steps in order, i. e. homography based LDV warping, energy optimization of overlap correspondences and energy optimization for hole filling. Then, the rectified LDV image is used to restore the missing pixels. We finally eliminate the ghost effect between the repaired region and its surroundhags by Poisson image blending.Experiments show that our method outperforns recent state-of-the-art image completion methods.