递进填充线算法对DIBR虚拟图像的修复

摘要

Depth-image-based rendering (DIBR) is an image synthesis method for rendering virtual free viewpoint images from depth images.However,virtual images generated by DIBR usually contain holes,which are spatial regions occluded by foreground objects.Many methods have been proposed to eliminate these holes.Criminisi and template matching are two of them.In this paper,a new hole-filling algorithm based on recursive-line is proposed to reduce the mismatches caused by the noise around contours of holes in Criminisi and template matching.First,according to the direction of view-switch,contours of holes in background area are widened to reduce the foreground noise from location deviation of depth image and view image.Second,in the ergodic hole-filling process,contours of holes in background area are extracted as filling lines.Only patches on filling lines can be filled,which reduces the number of candidate patches as well as gives high priority to background patches.Experimental results show the proposed algorithm gives higher PSNR than Criminisi algorithm by 1.617 3,and higher PSNR than depth-based template matching by 1.310 5.Proposed algorithm is capable of effectively reducing the errors in hole-filling and improving image quality.%在基于深度图的虚拟图像绘制(DIBR)中,前景物体的遮挡会造成虚拟图像中的空洞问题.针对Criminisi算法及模板匹配对空洞边界噪声敏感的问题,提出一种基于递进填充线的填补算法.首先根据视点变换的方向将空洞的背景部分边界扩大,以消除深度误差造成的空洞边界上的残余前景噪声;在空洞填充的遍历过程中,采用提出的递进填充线算法对每次更新后的剩余空洞提取基于视点方向的部分边界作为填充线,以限制被填充块的搜索范围,使被填充空洞块的择优顺序以背景纹理优先,避免了空洞边界上前景纹理的误填充.实验结果表明,该算法峰值信噪比比Criminisi算法平均高出1.6173,比基于深度的模板匹配算法平均高出1.3105,能更有效地抑制DIBR空洞纹理的错误生长,修复图像质量.