建筑立面图像重复结构检测方法研究

CASIA OpenIR > 毕业生 > 博士学位论文

	建筑立面图像重复结构检测方法研究
	肖鸿飞
	2017-12-04
学位类型	工学博士
中文摘要	建筑立面重复结构检测是建筑图像理解的重要任务。这一任务能够辅助众多计算机视觉问题，包括建筑物三维重建与建模、建筑图像编辑、城市场景理解等等。然而，这一问题具备以下三方面的难点。第一，对于不同风格建筑立面上的重复结构，它们的外观与布局通常存在着显著的差异。这限制了重复结构检测方法的通用性。第二，对于同一建筑立面的重复结构，由于窗帘收放等因素，它们的外观常存在着差别。第三，建筑立面图像常因为遮挡、倒影或变化光照等干扰因素而降质。由于以上难点，度量重复结构之间的相似度变得困难。为了应对上述难点，重复结构水平和竖直对齐的性质需要被利用。为此，本文提出了两种描述重复结构对齐性的模型，分别是矩阵乘法模型与基准线模型。结合重复结构的形状与颜色等特征，本文对重复结构检测提出了数种有效方法。具体地，论文的主要贡献包含以下四个方面： 1. 提出了基于显式矩阵分解的方法。该方法使用矩阵乘法模型描述重复结构对齐性。具体地，各重复结构被视作一个重复模式与两个分块矩阵的乘积。那两个分块矩阵是由交替的零矩阵与单位阵组成的，并且分别记录了重复结构的竖直水平位置。基于这一模型，检测重复结构转化为求解那两个分块矩阵的矩阵分解问题。为求解该矩阵分解问题，一种优化算法被提出，其中动态规划被用于优化分块矩阵。大量的实验结果表明了提出模型的有效性。 2. 提出了基于定位提取与优化的方法。这种方法的特点是无需任何标注数据而检测重复结构。该方法包含三步。第一，由于建筑结构通常是矩形的，实施矩形检测。第二，从检测矩形中，对应于真正建筑结构的矩形被提取。第三，基于已提取的矩形结构，一个优化问题被提出以寻找所有的重复结构。为求解该优化问题，一种高效的基于动态规划的算法也被设计。大量的实验结果证实了提出方法的有效性。 3. 提出了基于基准线提取的方法。因为重复结构是水平竖直对齐的，它们能够被经过重复结构边缘的水平竖直的基准线所定位。基于这一观察，该方法提出提取这些基准线，从而检测重复结构。首先，候选直线被检测，包含了所有的基准线与众多经过墙壁或重复结构的直线。然后，为了筛选基准线，该方法提出一个极大后验问题，用以度量筛选直线定位重复结构的概率。最后，该问题被一种基于动态规划的算法求解。大量的定性定量结果证实了提出方法的有效性。 4. 提出了基于分布距离最大化的方法。由于重复结构通常是矩形的，而且是水平竖直对齐的，本方法提出使用经过重复结构边缘的基准线约束重复结构的分割。重复结构分割问题被数学表述为一个约束优化问题。其中，通过最大化前景与背景的颜色分布之间的距离，基准线所决定的分割被优化。为求解该优化问题，一种基于动态规划的算法被提出。两个公开数据集上的实验结果证实了提出方法的可行性与有效性。
英文摘要	Facade repetition detection is an important task in facade image understanding. This task can facilitate many computer vision problems, including building modeling and 3D reconstruction, facade image analysis, urban scene understanding and so on. However, this problem has the following three difficulties. First, for repetitive structures on facades of different architectural styles, there is usually significant discrepancy among their appearances and layouts. This limits the generality of repetition detection approaches. Second, among the repetitions on a single facade, there are usually appearance variations caused by blind slides, shutter rotations and so on. Third, facade images are often degraded due to corruptions including occlusions, glass reflections and changing illuminations. Because of the above two difficulties, it becomes difficult to measure the similarities between the repetitions. To deal with the above difficulties, the property that facade repetitions are horizontally and vertically aligned requires to be utilized. For this purpose, this thesis proposes two models describing the alignment property, namely a matrix multiplication based model and fiducial lines passing along repetition boundaries. Combining with the shape, color and other features of the facade repetitions, this thesis proposes several effective approaches for repetition detection. Specifically, the main contributions of this thesis contain the following four aspects. 1. An explicit matrix factorization based approach is proposed. This approach utilizes a matrix multiplication based model to describe the alignment property among the facade repetitions. That is, the repetitions are viewed as the product of a repetitive pattern and two block matrices. The two block matrices are composed of alternating zero matrices and identity matrices, and record the vertical and horizontal positions of the repetitions respectively. Based on the model, repetition detection turns into a matrix factorization problem which optimizes the block matrices. An optimization algorithm is thus developed to solve the matrix factorization problem, where dynamic programming is used to optimize the block matrices. Extensive experiments demonstrate the effectiveness of the approach. 2. By localization, extraction and symmetry based optimization, an approach is proposed. This approach is characterised by detecting repetitions without requiring any labeled data. The approach contains three main modules. First, rectangle detection is conducted as facade structures are usually rectangular. Second, from the detected rectangles, the rectangles corresponding to real facade structures are extracted. Third, based on the extracted rectangular structures, an optimization problem is formulated to find all the repetitive structures. To solve the optimization problem, an efficient dynamic programming based algorithm is also developed. Comprehensive experimental results demonstrate the validity and effectiveness of the proposed approach. 3. An approach based on fiducial lines extraction is proposed. Since repetitions are horizontally and vertically aligned, they can be localized by the horizontal and vertical lines passing along the repetition boundaries. Based on the observation, the approach proposes to detect repetitions by extracting these fiducial lines. First, candidate lines are detected, containing both all the fiducial lines and many mistaken lines passing across facade wall or repetitive structures. Then, to pick out the fiducial lines, a maximum a posterior problem is formulated to measure the probabilities that the lines can localize the repetitions. Finally, the problem is efficiently solved by a dynamic programming based algorithm. Extensive qualitative and quantitative results verify the effectiveness of the approach. 4. An approach based on distribution distance maximization is proposed. Since the repetitions are generally rectangular, and horizontally and vertically aligned, this approach proposes to use fiducial lines, namely the horizontal and vertical lines passing along the repetition boundaries, to constrain the repetition segmentation. The problem of facade repetition segmentation is finally formulated as a constrained optimization problem, in which the segmentation determined by the fiducial lines is optimized by maximizing the distance between the foreground/background color distributions. An efficient dynamic programming based algorithm is also developed to solve the optimization problem. Experimental results on two publicly available datasets demonstrate the feasibility and validity of the proposed approach.
关键词	重复结构检测对称性检测建筑立面图像标注建筑立面图像解析
文献类型	学位论文
条目标识符	http://ir.ia.ac.cn/handle/173211/20649
专题	毕业生_博士学位论文
作者单位	中国科学院自动化研究所
推荐引用方式 GB/T 7714	肖鸿飞. 建筑立面图像重复结构检测方法研究[D]. 北京. 中国科学院研究生院,2017.

条目包含的文件
文件名称/大小	文献类型	版本类型	开放类型	使用许可
ThesisV5.pdf（5837KB）	学位论文		限制开放	CC BY-NC-SA