基于多尺度超像素分割的立体匹配算法研究

发布时间：2018-06-15 04:31

  本文选题：双目视觉 + 立体匹配　； 参考：《浙江大学》2017年硕士论文

【摘要】：立体视觉及三维重建技术日益受到各行业的关注和重视,尤其是虚拟现实、增强现实及混合现实技术发展得如火如茶,立体匹配作为立体视差法三维重建技术的关键有着重要的研究价值。目前从基本的局部SAD及NCC相似性度量到全局能量的置信传播及图割方法衍生的立体匹配算法层出不穷,但对于遮挡问题及非子模能量函数的优化仍是一项重大的挑战。本文采用3D标签来描述各个像素的空间特征,在已有的马尔可夫随机场最大后验(MRF-MAP)能量代价模型下,采用3D标签的法向及切平面特征来建立二阶先验的立体曲面。二阶先验平滑项放松了前向平行平面的约束使得满足斜向切平面先验时也不受到惩罚,使得水平及垂直方向邻域的各像素可以相互关联。针对遮挡及多重团簇的非子模性难题,本文改造了非对称图模型来控制团簇大小,通过增加可见性节点将所有可能的遮挡信息呈现在图中,并将可见性节点的关联边整合到能量函数的数据项中。同时本文改进了伪布尔函数多项式优化(QPBO),为每个像素节点创立与之对立的非节点,使得非子模的边也可用mincut/max flow方法进行分割,改进的QPBOI-R方案能在不增加能量代价的情况下合理地更新不可标记节点,经过多次迭代可得到近似最小代价的解。本文最主要的贡献是提出了多尺度超像素候选标签图(Multi-scale Super-pixel basedProposals,MSP)结构及其更新方法,已有的候选图更新方法有基于图像分割的或随机生成的。MSP通过随机选取已修正的标签来更新超像素图像的候选标签,主要是利用了多个尺度的超像素图像可以在多种特征层面下更新标签而不影响边缘处的视差突变。另外,将多个超像素候选与两个棋盘格候选相结合构成完整的候选标签图,超像素候选表征着区域的特征信息而棋盘格保留了像素级别的特征信息,实验证明了 MSP结构下的视差图在均匀无纹理区域呈现平滑特性,而在边缘处又呈现了深度不连续特性。
[Abstract]:Stereo vision and 3D reconstruction technology have attracted more and more attention from various industries, especially virtual reality, augmented reality and mixed reality technology. Stereo matching plays an important role in 3D reconstruction of parallax. At present, stereo matching algorithms derived from basic local SAD and NCC similarity measures to global energy confidence propagation and graph cutting methods are emerging in endlessly. However, it is still a major challenge for occlusion problems and optimization of nonsubmode energy functions. In this paper, 3D tags are used to describe the spatial features of each pixel. Under the existing Markov random field maximum posterior MRF-MAP-based energy cost model, the normal and tangent plane features of 3D tags are used to establish the second-order priori stereoscopic surfaces. The second-order priori smoothing term relaxes the constraint of the forward parallel plane so that the apriori of oblique tangent plane is not penalized so that the pixels in the horizontal and vertical direction neighborhood can be correlated with each other. In order to solve the non-submodule problem of occlusion and multiple clusters, the asymmetric graph model is modified to control the cluster size, and all possible occlusion information is presented in the graph by adding visibility nodes. The associated edges of the visibility nodes are integrated into the data items of the energy function. At the same time, we improve the pseudo-Boolean function polynomial optimization and create a non-node for each pixel node, so that the edges of non-submodules can also be segmented by mincut/max flow method. The improved QPBOI-R scheme can reasonably update the unlabeled nodes without increasing the energy cost. After several iterations, the approximate minimum cost solution can be obtained. The main contribution of this paper is to propose the multi-scale super-pixel based proposals MSPs structure and its updating method. The existing candidate image updating methods include. MSP, which is based on image segmentation or randomly generated, updates candidate labels of superpixel images by randomly selecting modified labels. The multi-scale super-pixel image can update the label at multiple feature levels without affecting the parallax mutation at the edge. In addition, a plurality of super-pixel candidates are combined with two checkerboard candidates to form a complete candidate tag map. The super-pixel candidate represents the feature information of the region, while the checkerboard retains the pixel level feature information. The experimental results show that the parallax image with MSP structure is smooth in the homogeneous texture-free region, and the depth discontinuity is shown at the edge of the parallax image.
【学位授予单位】：浙江大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TP391.41
，

本文编号：2020668