基于GPU加速的虚拟视点合成研究与实现

发布时间：2019-02-20 07:30

【摘要】：自由视点电视是三维立体电视的一个分支,是未来电视发展的方向。相比传统的3D电视,自由视点电视能够带给人们更好的立体体验。虚拟视点合成技术是自由视点电视的核心技术,其中最受关注的是基于深度图的绘制技术(Depth Image Based Rendering,DIBR)。利用DIBR技术可以快速生成虚拟视点图像,但是仅仅依靠DfIBR技术合成的图像并不理想,往往包含一些“伪影”和大小不等的空洞。而且,现有的基于DIBR技术的虚拟视点合成算法一般比较耗时,很难做到虚拟视点的实时合成。本文主要的研究内容是:分析以上问题产生的原因,并有针对性的提出相应的解决方案。(1)“伪影”往往出现在前景和背景相交的地方,主要由深度的不连续性和多个相机拍摄到的场景的颜色差异造成。针对“伪影”问题,本文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,来减弱“伪影”现象。(2)空洞问题是虚拟视点合成中最重要,也最难解决的问题,主要是由视点变换中物体间遮挡关系发生变化造成的。针对较小的空洞,本文主要采用深度图预处理和反向映射方法;针对较大的空洞,本文采用了基于辅助深度信息的图像修复算法,但是该算法运行比较耗时。为了加快空洞填补速度,本文将基于辅助深度信息图像修复算法并行化,通过 NVIDIA 公司推出的 CUDA(Compute Unified Device Architecture)并行编程框架,利用GPU来加速该算法,可以大大减少空洞填补时间。实验结果表明,本文针对“伪影”问题的处理及基于辅助深度信息的图像修复的空.洞填补算法可以有效地提高虚拟视点图像合成质量,同时基于CUDA实现的并行空洞填补算法也大幅降低了空洞填补时间。
[Abstract]:Free-view TV is a branch of three-dimensional TV and the direction of TV development in the future. Compared with traditional 3D TV, free view TV can bring people a better stereoscopic experience. Virtual view synthesis technology is the core technology of free view television, and the most concerned technology is depth map based rendering technology (Depth Image Based Rendering,DIBR). The virtual view image can be generated quickly by using DIBR technology, but the image synthesized only by DfIBR technology is not ideal, and often contains some "artifacts" and holes of different sizes. Moreover, the existing virtual viewpoint synthesis algorithms based on DIBR technology are usually time-consuming, and it is difficult to achieve real-time virtual view synthesis. The main research contents of this paper are as follows: analyzing the causes of the above problems and putting forward corresponding solutions. (1) "artifact" often appears in the place where the foreground and background intersect. It is mainly caused by the discontinuity of depth and the color difference of the scene captured by multiple cameras. Aiming at the problem of "artifact", the color correction of reference view image is carried out in this paper, and the color image fusion based on depth and the color image fusion directly are integrated. (2) void problem is the most important and most difficult problem in virtual view synthesis, which is mainly caused by the change of occlusion relationship between objects in view transformation. For the smaller voids, this paper mainly uses depth map preprocessing and reverse mapping methods, for the larger holes, this paper uses an image restoration algorithm based on auxiliary depth information, but the algorithm runs more time-consuming. In order to speed up the filling speed of holes, this paper parallelizes the image restoration algorithm based on auxiliary depth information. Through the parallel programming framework of CUDA (Compute Unified Device Architecture) developed by NVIDIA Company, using GPU to accelerate the algorithm, the time of filling holes can be greatly reduced. The experimental results show that the processing of "artifact" and the space of image restoration based on auxiliary depth information are discussed in this paper. The hole filling algorithm can effectively improve the quality of virtual view image synthesis, and the parallel hole filling algorithm based on CUDA also reduces the void filling time significantly.
【学位授予单位】：东南大学
【学位级别】：硕士
【学位授予年份】：2016
【分类号】：TP391.41

【参考文献】