>> 徐胜华，2007，面向立体影像特征匹配的直线提取方法
摘 要
随着经济的发展和城市化进程的加快，获取地球表面高分辨率、高精度、多时相、多光谱数字影像的技术已日趋成熟并得到广泛应用。基于高分辨率遥感影像提取三维地球空间信息对于城市规划、环境监测、交通管理，城市GIS 数据快速更新等具有十分重要的意义，特别是数字摄影测量技术的发展为获取现实世界中人工物体如建筑物的三维空间信息提供了极大的方便，直线特征提取是数字摄影测量三维城市建模的关键科学问题之一，同时也是计算机视觉和图像理解领域的研究前沿。鉴于线特征具有比点特征更丰富的信息且比面特征更容易提取和描述，本文把线特征提取作为研究的重点。然而，由于直线断裂、遮挡以及共面空间直线投影等因素的影响，从左右影像中提取出来的直线之间会出现“一配多”甚至“多配多”的匹配情况，可靠地处理这些问题成为三维自动化重建的瓶颈之一。本文研究面向立体影像特征匹配的直线提取方法，对提取的直线进行融合和筛选，通过直线编组，每个特征编组包含了内部直线之间的匹配关系。直线匹配的问题因此转化为从直线集中提取一些相互兼容的特征编组问题，大大降低了直线匹配的难度。具体内容包括：
(1) 改进了彩色图像边缘提取与亚像素定位方法。边缘提取是获取图像特征的基本方法，在回顾现有的主要彩色图像边缘检测方法原理的基础上，分析了边缘检测的难点及存在的主要问题，提出了一种综合运用主轴分析法、Zernike矩和嵌入置信度的亚像素边缘检测算法。首先获取影像三维空间中的主轴，然后利用嵌入置信度的边缘检测算子快速检测出所有可能的边缘点，再利用Zernike 矩算子以亚像素精度重新定位边缘。实验表明本文算法优于其它的边缘检测算子，定位精度能够达到0.24 个像素，并且该方法还具有良好的抗噪声干扰性和处理效率。
(2) 基于Hough 变换的直线提取方法。针对Hough 变换提取直线算法在速度、精度和影像大小三方面的局限，提出了一种基于预存储权值矩阵的多尺度Hough 变换算法。首先阐述了预存储权值矩阵的Hough 变换原理，以及权值矩阵的计算和存储特点，对参数空间中ρ 的分辨率和θ 的分辨率的最佳取值作了探讨，之后详细说明了基于预存储权值矩阵的多尺度Hough 变换直线提取算法。实验证明，本文算法显著提高了实际影像处理的速度和直线提取的精度，特别是对比较大的影像具有计算量小、抗噪能力强等特点。
(3) 基于D-S 证据理论的彩色航空影像阴影提取方法。阴影特征在影像，特别是遥感影像中具有重要的意义。针对彩色航空遥感影像的颜色信息进行分析，提出了表述阴影在RGB 和HIS 空间的三种不同特征，并利用均值平移(MeanShift)分割的方法对彩色航空影像进行区域分割。选取分割后的区域，统计每个区域的多个颜色特征并定义相应的概率分配函数，利用其互补特性，使用Dempster-Shafer 证据理论中的合成法则对其进行合成，最终判别区域是阴影还是非阴影。
(4) 基于感知编组的直线融合和筛选策略。首先简要介绍了认知心理学的基本概念和感知编组理论。提出了一种利用K-Means 分类、阴影信息和Mean Shift区域分割的感兴趣区域提取方法。然后利用感兴趣区域信息对边缘和相位编组法提取的直线进行筛选，之后对由两种不同方法提取的直线进行融合和筛选。充分利用人工地物的结构信息进行编组来获取直线间的语意关系，使得编组后的直线主要由人工地物的轮廓组成，包括长直线段、“L”型结构、“U”型结构、平行直线和多边形。
本文所研究提出的面向立体影像特征匹配的直线提取的有关理论、方法与算法已用VC++6.0 编程实现。有关实验结果证明本文的直线提取算法能有效提取立体影像对中的直线特征，为后续基于直线特征的立体匹配方法研究打下了坚实的基础。总之，直线特征提取在今后相当长的时期内仍将是数字摄影测量和计算机视觉领域内的一个难题，深入开展直线特征提取及其相关领域技术的研究对于进一步推广数字摄影测量系统的实际应用具有重要的研究意义。
关键词：
立体影像特征匹配，亚像素边缘检测，直线提取，阴影提取，感知编组
Abstract
With the rapid development of economy and urbanization, technologies of acquiring
digital images with high-resolution and multi-spectra are becoming more and more mature
and have gotten increasing applications. 3D geospatial information is very important to
various fields, such as city planning, environment monitoring, traffic controlling and
updation of GIS data. Especially, the development of digital photogrammetry greatly
facilitates 3D geospatial information extraction of the real world objects to a great extent.
But the theory and technology of data processing and analysis based on images lag behind
correspondingly and feature extraction based on images is the main barrier. How to extract
straight line information for 3D reconstruction of man-made objects from images has
always been concerned by the researcher from the fields of photogrammetry and computer
vision.
Linear feature can provide more information than point feature and can be extracted
and described conveniently compared with the planar structure. So Line extraction is the
key issue of this dissertation. But due to line fragmentation, occlusion and projection of
conjoint coplanar space straight lines, there are many “one-to-multiple” and even
“multiple-to-multiple” mappings between two features sets in the process of stereoimage
matching, to deal with such kind of problems reliably is the bottleneck of image
processing. Therefore, a new straight line extraction method for stereoimage feature
matching is proposed. First, the fusion and the filtering strategies are designed for the
linear features extracted from the images, then feature grouping is implemented among the
linear features, and each feature group contains its associated matching relationships. In
this way, stereo matching becomes equivalent to extracting a set of mutually compatible
feature groups from the image. The main work of this dissertation can be summarized as
follows:
(1) An improved color image edge detecting algorithm and sub-pixel location
method. Edge detection is the base of the image measuring, image segmentation,
image compressing and pattern recognition and so on. Based on the analysis of
the current means of color image edge detection, the main problems and
difficulties in edge detecting are discussed. This dissertation discusses a new
approach for detecting edges with sub-pixel accuracy in color images using the
principal axis analysis, Zernike moments and embedded confidence. First, the
principal axis of the color vectors in a three-dimensional space is obtained
analytically for the image. Then all the possible edge points are quickly
measured by using the edge detection with embedded confidence. At last the
edge is relocated with sub-pixel accuracy by means of Zernike moment operator.
Experiments show that the accuracy of the proposed method is superior to other
edge detectors. The location accuracy of 0.24 pixels is obtained. Results from
intensive experimental work also indicate that this method has the desired
performance of anti-interference and is highly efficient.
(2) Line extraction method based on Hough Transform. Aiming at the critical
time-consuming and accuracy issues of straight line extraction from large-scale
remote sensing imagery, after briefly reviewing the existing straight line
extraction methods, a multi-scale Hough transform method based on the
pre-storage weight matrix is proposed, which saves a lot of storage space, takes
care of discretization errors, and avoids the abruption and conglutination of
characters that are the drawbacks of the existing straight line extraction
algorithms. This dissertation introduces the principle and the implementation
algorithm of multi-scale Hough transform method based on the pre-storage
weight matrix in detail. To optimize speed and precision, the best choice of
accumulator based on image size is suggested too. The experimental results show
that this algorithm is more efficient in computation and robust to noise, and is
rich in feature content and accurate, especially for large-size images.
(3) A new shadow extraction method from color aerial image based on
Dempster-Shafer (D-S) evidence theory. Shadows play an important role in our
understanding of imagery especially the remote sensing imagery. According to
the color features of shadow in the color aerial images, a new approach is
proposed to extract shadow from color aerial images based on the D-S evidence
theory. Firstly three different feature descriptions of shadow in RGB and HIS
spaces are introduced, then the image is segmented based on Mean Shift method,
the color features of the segmented regions are computed and the basic
probability assignment function (BPAF) of each segment is defined respectively.
By using of the complementary characteristics of such three color features, the
D-S rule is applied to compute the fusion of BPAF. Finally, the BPAF fusion is
utilized to find the conclusive shadow segments.
(4) The fusion and the filtering strategies based on perceptual grouping. The Gestalt
theory in cognitive psychology and the perceptual grouping theory are briefly
introduced. A method of extracting regions of interest (ROI) is proposed using
the K-Means classification, shadow information and Mean Shift segmentation.
The information of ROI is used to filter the edges and the lines extracted by
phase-grouping method. Then the lines extracted by two different algorithms are
fused and filtered utilizing basic rules of perceptual grouping. The semantic
interrelationships between linear features are exploited by perceptual grouping to
extract structure to detect the presence of manmade objects. These features are
generated by the strong boundaries typical of manmade structures: longer linear
lines, “L” junctions, “U” junctions, parallel lines, and polygons.
The software prototype of straight line extraction for stereoimage feature matching
proposed in this dissertation is implemented using VC++6.0. From the experimental
results, this method is preferable to extract linear features from aerial stereo pairs, which
will lay the solid foundation for the future research on the stereoimage matching method
based on line feature.
In conclusion, the straight line extraction will be one of the difficult questions in the
field of photogrammetry and computer vision in a very long period. It is very important
for the practical use of digital photogrammetry system to develop more reliable line
extraction techniques.
Key words: Stereoimage feature matching, Edge detection with sub-pixel precision, Line
extraction, Shadow extraction, Perceptual grouping.
