CASIA OpenIR  > 毕业生  > 硕士学位论文
初期视觉中的几个问题及张量可视化研究
其他题名STUDIES ON SOME PROBLEMS IN EARLY VISION AND TENSOR VISUALIZATION
周渊
学位类型工学硕士
导师李泉林 ; 郑应平 ; 胡占义
2001-05-01
学位授予单位中国科学院自动化研究所
学位授予地点中国科学院自动化研究所
学位专业模式识别与智能系统
关键词计算机视觉 边缘检测滤波器 随机hough变换(Rht) 张量场 可视化 Ph分布 Computer Vision Edge Detection Filter Randomized Hough Transform (Rht) Tensor Field Visualization Ph Distribution
摘要边缘检测是研究计算机视觉的基本问题。几何基元提取是任何计算机视觉系 统的关键组成部分和基本要求。 本文主要研究的是Marr视觉计算理论的第一阶段(也称为初期视觉、早期 视觉)中的边缘检测滤波器和Hough变换算法问题。所围绕主线为:为了准确 地提取出几何基元,首先对原始图进行滤波、边缘检测等前期处理,从灰度图中 得到只包含特征点和背景点的二值图像。然后,运用随机Hough变换提出几何基 元。除此之外,本人还研究了计算机图形学中科学可视化问题,完成了 Schlumberger石油公司与中科院自动化所的一个合作项目。本文所做的主要工作 可以归纳为以下四个方面: 1.对称位相型滤波器及其在边缘检测中的应用。 从概率的角度提出了两类新的滤波器:对称位相型滤波器和对称梯阶滤波 器;证明了对任意给定的非负对称滤波器,可用这两类滤波器的序列在(-∞,+∞) 内一致逼近;这种序列有着统一的结构形式: 对于L(x),我们给出了选择L(x)一些最优准则的解析表达式,从而表明 L(x)既具有一般性,又易进行解析处理和数值计算。对于已有文献中相关于对 称指数滤波器的两类不同形式,例如Li和Ma[1994]和Ma和Li[1995],我们 从理论上进一步地刻画了它们的内在一致性。 2.对称位相型x2-滤波器及其在边缘检测中的应用。 利用位相型分布(见Neuts[1981])提出了对称位相型滤波器 L(x)=gaexp{Tx2}T0, 其中是一个正则化正常数;证明了对 任意给定的非负对称滤波器,可用这类滤波器的序列在(-∞,+∞)内一致逼近; 这种序列有着统一的结构形式: 对于L(x),我们可以给出选择L(x)一些最优准则的解析表达式。我们的结 果将文献中较为重要的高斯滤波器,高斯一拉普拉斯两阶微分滤波器(见 Marr[1982]和Canny[1983])以及两个尺度比为t的高斯滤波器之差(见Ma and Li[1995])等形式在一致逼近的意义下推进到较为一般的情况,并且实现了解析 处理和数值分析。 3.随机Hough变换的概率模型 研究了基于有限个数据点的随机Hough变换的概率模型。在这个模型中, 我们主要讨论了在随机.Hough变换的基本算法中起相当关键作用的两个量:累 加器数组的控制阈值,和从图像中提取全部基元所需随机抽样的总次数。这两个 量对随机Hough变换的算法设计及其终止规则的确定是相当有用的。 主要结果包括两部分:其一是我们对累加器数组引入多项分布,系统地研究 了累加器数组的概率结构及其相互关系。同时我们也
其他摘要Edge detection is the basic problem of computer vision. And geometric primitive extraction is the element component and the fundamental requirement of any computer vision systems. In this work, we have mainly studied the edge detection filter and Hough transform in the first stage of the Marr Vision Theory (that is also called the Early Vision). The main line is: In order to extract geometric primitive precisely, the real picture is filtered and edge-detected to get the binary picture which only contains feature points and background points; then the Randomized Hough Transform (RHT) is applied to find the primary sketch. Besides, I have also researched in the scientific visualization, and have finished the Geostress tensor visualization, a cooperation project between the Schlumberger Oil Company and Institute of Automation, CAS. In short, the following four topics are investigated in this thesis: 1. Symmetric Phase Type Filters and Their Applications to Edge Detections We present two classes of new filters from the viewpoint of probability: the symmetric phase type filter and the symmetric step type filter. We also prove that the sequences of the two classes of filters can uniformly approximate any arbitrarily given non-negative symmetric filters within (-∞,+∞). The sequences are shown to have a unified form The analytic expressions of some optimal rules are given for selecting the optimal filter L(x). Thus we demonstrate that L(x) is not only general but also easy to analytically and numerically consider. For example, Li and Ma [ 1994] and Ma and Li [ 1995], the two papers is shown to have an intrinsic uniformity from our results. 2. Symmetric Phase Type x2 -Filters and Their Applications to Edge Detections We present one class of new filters from the viewpoint of phase type distributions: the symmetric phase type filter. We also prove that the sequences of the filters can uniformly approximate any arbitrarily given non-negative symmetric filters within (-∞,+∞). The sequences are shown to have a unified form The analytic expressions of some optimal rules are given for selecting the optimal filter L(x). Thus we demonstrate that L(x) is not only general but also easy to analytically and numerically consider. For example, Marr[1982], Canny[1983] and Ma and Li [1995], the different filters are shown to have an intrinsic uniformity from our results. 3. Probability Model of the Randomized Hough Transform (RHT) We provide a probability model of randomized Hough transform based on limited data. In this model, we mainly discuss the two key quantities: the control threshold value of the accumulator arrays and the total number of random sampling for extracting all the primitives from the image. The two quantities are quite useful for to how to design some effective algorithms of randomized Hough transform and how to determine their stopping rules. The main re
馆藏号XWLW596
其他标识符596
语种中文
文献类型学位论文
条目标识符http://ir.ia.ac.cn/handle/173211/7331
专题毕业生_硕士学位论文
推荐引用方式
GB/T 7714
周渊. 初期视觉中的几个问题及张量可视化研究[D]. 中国科学院自动化研究所. 中国科学院自动化研究所,2001.
条目包含的文件
条目无相关文件。
个性服务
推荐该条目
保存到收藏夹
查看访问统计
导出为Endnote文件
谷歌学术
谷歌学术中相似的文章
[周渊]的文章
百度学术
百度学术中相似的文章
[周渊]的文章
必应学术
必应学术中相似的文章
[周渊]的文章
相关权益政策
暂无数据
收藏/分享
所有评论 (0)
暂无评论
 

除非特别说明,本系统中所有内容都受版权保护,并保留所有权利。