欢迎访问《图学学报》 分享到：
图学学报

图学学报

• 视觉与图像 • 上一篇    下一篇

一种快速消除失真的虚拟视点合成方法

  

  1. 1. 东南大学计算机科学与工程学院、软件学院，江苏 南京 211189；
    2. 计算机软件新技术国家重点实验室(南京大学)，江苏 南京 210023
  • 出版日期:2017-08-31 发布日期:2017-08-10
  • 基金资助:
    国家高技术研究发展计划(863计划)项目(2015AA015904)；NSFC-广东联合基金(第二期)超级计算科学应用研究专项

Virtual Viewpoint Synthesis of Quick Remove the Distortion

  1. 1. School of Computer Science and Engineering, College of Software Engineering, Southeast University, Nanjing Jiangsu 211189, China;
    2. State Key Laboratory for Novel Software Technology, Nanjing University, Nanjing Jiangsu 210023, China
  • Online:2017-08-31 Published:2017-08-10

PDF (PC)

1452

摘要： 基于深度图的绘制(DIBR)技术是合成虚拟视点图像的有效方法，但是合成的图像
往往存在失真，最主要的问题是包含“伪影”和大小不等的空洞。针对“伪影”问题，先对参考视
点图像进行色彩校正，并综合基于深度的彩色图融合和彩色图直接融合两种方式融合图像，以
减弱“伪影”效应；针对映射后出现的小空洞，采用反向映射到参考图像的方式进行填充；针对
大的空洞，采用图像修复的方式，结合深度图更好地填补空洞，但是图像修复是一个耗时的过
程，为此将图像修复算法并行化，利用统一计算设备架构(CUDA)并行计算，来加速空洞填补过
程。实验结果表明，针对“伪影”问题的处理及基于图像修复的空洞填补可以有效地提高虚拟视
点图像合成质量，同时基于CUDA 实现的并行图像修复算法也大幅降低了空洞填补时间。

关键词: 自由视点视频, 视点合成, 图像修复, GPU 加速

Abstract: Depth image based rendering (DIBR) is an effective method for virtual view synthesis. But
synthetic images often contain distortion, the main problem is ghost effect and some holes of varying
sizes. To cope with the ghost effect, the color correction of reference views is first used, and then two
fusion methods are combined which are depth-based color image fusion and direct color image fusion.
Inverse warping is adopted to fill cracks. In order to fill the remaining disocclusions, the image
inpainting with depth map is used to get a better virtual view. Because the process of image inpainting
is very time-consuming, the inpainting algorithm on GPU is implemented to accelerate the inpainting
process based on compute unified device architecture (CUDA). Experimental results show that our
proposed method improved the quality of virtual view synthetic images, and the image inpainting
algorithm based on CUDA implemented also significantly reduced the time of hole filling.

Key words: free-view video, view synthesis, image inpainting, GPU acceleration