基于视差估计与校正的虚拟视点绘制技术研究

发布时间：2018-05-26 22:12

本文选题：视差估计 + 校正　；参考：《福州大学》2014年硕士论文

【摘要】：随着三维显示技术的发展,三维电视系统(3DTV)也逐步的完善,即将向下个目标自由立体显示迈进。在自由立体显示系统中,用户能自由的切换场景中的视角,拥有更好的用户体验。然而,我们不可能使用无限多个摄像机进行场景的采集,因此,目前主要采用虚拟视点绘制技术,通过场景中有限的视角,产生出其他视角的影像。但是,虚拟视点的绘制技术还存在着不少问题需要解决。因此,一种高效合理的虚拟视点绘制技术是目前三维立体显示中研究的热点。本文主要是针对目前虚拟视点绘制中DIBR技术出现的问题进行改进,包括深度图获取过程中,在场景不连续和低纹理区域中产生的误差,以及在虚拟视点绘制中空洞填补出现的重影和不协调等现象进行处理。论文研究的主要内容包括以下部分：首先研究了区域立体匹配算法中,场景不连续区域和低纹理区域出现的误匹配问题,提出一种可变窗口的视差估计方法,该方法是利用较大的窗口先进行区域匹配,选择可靠性较高的匹配点,之后采用mean-shift图像分割算法将图像分割,利用分割后的图像平滑约束条件,将可靠的匹配点扩散到整个区域,从而,提高整体的视差估计精度,算法相比传统的自适应匹配算法,视差不连续区域匹配误差降低10%左右。其次,将立体匹配得到的视差图转换成深度图,进行基于深度图的虚拟视点绘制,针对双向虚拟绘制技术中的空洞填补,由深度图不精确造成的误差,提出一种深度误差校正的空洞填补技术,首先,分别对左右参考图像进行三维映射,提取参考图像空洞的邻域与目标图像进行SAD匹配,计算匹配代价最小的偏移值。然后将目标图像按照获取的偏移量进行调整,从而,对深度图产生的误差进行校正,提高空洞填补的精度。最后,针对双向虚拟视点绘制中出现的伪影和场景不协调现象,采用一种线性融合技术,将空洞的邻域进行线性融合。在目标图像经过SAD校正后,根据邻域像素的可靠性,由参考图和目标图按照一定的比例进行融合,因此,该方法不会出现重影现象,而且可以减轻场景的伪影和不协调等现象。相比传统的双向DIBR算法,本文算法的PSNR值提升0.77dB左右。虚拟视点绘制技术作为自由立体显示系统中关键部分。本文针对传统的DIBR算法存在的问题,进行分析和改善。在主观视觉观看效果和客观评价上都有一定的提升。
[Abstract]:With the development of 3D display technology, 3D TV system (3DTV) is gradually improved, and the next target is Auto-stereoscopic display. In the stereoscopic display system, the user can freely switch the perspective of the scene and have a better user experience. However, it is impossible for us to use infinite cameras to capture the scene. Therefore, at present, we mainly use virtual view rendering technology to produce images of other visual angles through the limited visual angle in the scene. However, there are still many problems to be solved in virtual view rendering technology. Therefore, an efficient and reasonable virtual view rendering technology is a hot topic in 3D stereo display. This paper mainly aims at improving the problems of DIBR technology in virtual view rendering, including the errors in scene discontinuity and low texture region in depth map acquisition. And in the virtual point of view rendering holes filled in the appearance of double-shadow and disharmony and other phenomena to deal with. The main contents of this paper are as follows: firstly, the mismatch problem in the discontinuous region and low texture region in the region stereo matching algorithm is studied, and a variable window disparity estimation method is proposed. The method uses large window to match the region first, selects the matching point with high reliability, then uses mean-shift image segmentation algorithm to segment the image, and uses the constraint condition of image smoothing after segmentation. The reliable matching points are spread to the whole region, so that the overall disparity estimation accuracy is improved. Compared with the traditional adaptive matching algorithm, the disparity discontinuous region matching error is reduced by about 10%. Secondly, the parallax map obtained by stereo matching is converted into depth map, and the virtual viewpoint drawing based on depth map is carried out. The errors caused by the inaccuracy of depth map are filled in the void of bidirectional virtual rendering technology. A hole filling technique for depth error correction is proposed. Firstly, 3D mapping of left and right reference images is carried out, and the neighborhood of the reference image is extracted for SAD matching with the target image, and the offsets with the least matching cost are calculated. Then the target image is adjusted according to the obtained offset, and the error caused by the depth map is corrected to improve the accuracy of cavity filling. Finally, aiming at the artifacts and scene disharmony in bidirectional virtual view rendering, a linear fusion technique is used to fuse the empty neighborhood. After the target image is corrected by SAD, according to the reliability of the neighborhood pixels, the reference map and the target map are fused according to a certain proportion. Therefore, the method does not appear the phenomenon of double shadow, and can reduce the phenomenon of artifact and incongruity of the scene. Compared with the traditional bidirectional DIBR algorithm, the PSNR value of this algorithm is increased by about 0.77dB. Virtual view rendering technology is a key part of the autostereoscopic display system. This paper analyzes and improves the problems existing in the traditional DIBR algorithm. In subjective visual viewing effect and objective evaluation have a certain improvement.
【学位授予单位】：福州大学
【学位级别】：硕士
【学位授予年份】：2014
【分类号】：TN873

【相似文献】