Robust depth estimation for light field via spinning parallelogram operator
Robust depth estimation for light field via spinning parallelogram operator
复制标题
DOI：
10.1016/j.cviu.2015.12.007
复制
发表时间：
2016-04-01
影响因子：
4.5
通讯作者：
Xiong, Zhang
中科院分区：
文献类型：
--
作者：
Zhang, Shuo;Sheng, Hao;Xiong, Zhang
去除遮挡对光场图像深度估计的影响一直是一个难题，特别是对于全光相机捕获的高噪声和混叠图像。在本文中，一个旋转的搜索算子（SPO）集成到一个深度估计框架来解决这些问题。利用算子在对极平面图像（EPI）中划分的区域，通过最大化区域的分布距离来定位指示深度信息的线。与传统的多视图立体匹配方法不同，即使被遮挡或有噪音，距离度量也能够保持正确的深度信息。我们进一步在光场丰富的结构中选择相对可靠的信息，以减少遮挡和模糊的影响。离散标记问题，然后解决了基于滤波器的算法，以快速逼近最优解。所提出的方法的主要优点是，它是不敏感的遮挡，噪声和混叠，并没有要求的深度范围和角分辨率。因此，它可以用于各种光场图像，特别是全光相机图像。实验结果表明，该方法在光场图像（包括真实的世界图像和合成图像）上的深度估计性能优于现有的深度估计方法，尤其是在遮挡边界附近。(C)2016 Elsevier Inc. All rights reserved.
Removing the influence of occlusion on the depth estimation for light field images has always been a difficult problem, especially for highly noisy and aliased images captured by plenoptic cameras. In this paper, a spinning parallelogram operator (SPO) is integrated into a depth estimation framework to solve these problems. Utilizing the regions divided by the operator in an Epipolar Plane Image (EPI), the lines that indicate depth information are located by maximizing the distribution distances of the regions. Unlike traditional multi-view stereo matching methods, the distance measure is able to keep the correct depth information even if they are occluded or noisy. We further choose the relative reliable information among the rich structures in the light field to reduce the influences of occlusion and ambiguity. The discrete labeling problem is then solved by a filter-based algorithm to fast approximate the optimal solution. The major advantage of the proposed method is that it is insensitive to occlusion, noise, and aliasing, and has no requirement for depth range and angular resolution. It therefore can be used in various light field images, especially in plenoptic camera images. Experimental results demonstrate that the proposed method outperforms state-of-the-art depth estimation methods on light field images, including both real world images and synthetic images, especially near occlusion boundaries. (C) 2016 Elsevier Inc. All rights reserved.