没有合适的资源?快使用搜索试试~ 我知道了~
温馨提示
The sources of error for the edge finding technique proposed by Marr and Hildreth (D. Marr and T. Poggio, Proc. R. Soc. London Ser. B204, 1979, 301–328; D. Marr and E. Hildreth, Proc. R. Soc. London Ser. B.207, 1980, 187–217) are identified, and the magnitudes of the errors are estimated, based on idealized models of the most common error producing situations. Errors are shown to be small for linear illuminations, as well as for nonlinear illuminations with a second derivative less than a critical value. Nonlinear illuminations are shown to lead to spurious contours under some conditions, and some fast techniques for discarding such contours are suggested.
资源推荐
资源详情
资源评论
COMPUTER VISION, GRAPHICS, AND IMAGE PROCESSING 27, 195-210 (1984)
Accuracy of Laplacian Edge Detectors
VALDIS BERZINS
Computer Science Department, University of Minnesota, Minneapolis, Minnesota 55455
Received September 3, 1983; revised November 9, 1983
The sources of error for the edge finding technique proposed by Marr and Hildreth (D. Marr
and T. Poggio,
Proc. R. Soc. London Ser.
B 204, 1979, 301-328; D. Marr and E. t-Iildreth,
Proc. R. Soc. London Ser. B. 207,
1980, 187-217) are identified, and the magnitudes of the
errors are estimated, based on idealized models of the most common error producing situa-
tions. Errors are shown to be small for linear illuminations, as well as for nonlinear illumina-
tions with a second derivative less than a critical value. Nonlinear illuminations are shown to
lead to spurious contours under some conditions, and some fast techniques for discarding such
contours are suggested.
1. INTRODUCTION
An important problem in low level image analysis is the detection of edges, which
can then be used to identify objects in a scene. In their work on scene analysis [1, 2],
Marr and Hildreth proposed an important technique for detecting edges in two
dimensional digitized images. In this paper we analyze the accuracy of this tech-
nique. Surveys of other edge detection algorithms can be found in [3-5], and
analyses of the accuracy of other edge detectors can be found in [6, 7].
Marr and Hildreth selected a limited range of spatial frequencies by blurring the
image with a Gaussian filter, and identified edges as the locus of points where the
directional derivative of the filtered image has a peak, which implies the second
directional derivative has a zero crossing. Their technique is important because its
ability to locate edges at several different resolutions, determined by the standard
deviations of the Gaussian filters, is useful for solving the matching problem in
stereo vision [8]. Marr and Hildreth showed that in cases where the variation of the
smoothed image parallel to the edge is at most linear, the value of the Laplacian is
the same as that of the second directional derivative in the direction in which the
slope of the zero crossing is steepest, which is at fight angles to the locus of zero
crossings. These results are important because the Laplacian is an orientation-inde-
pendent operator. They also showed that the Gaussian filter was optimal in the sense
that the response of the filter is as narrow as possible in both the spatial and the
frequency domain, thus minimizing the effects of aliasing introduced by a band
limited filter. The technique is amenable to approximations that allow efficient
calculation of edge contours, and it appears to do well in empirical tests, making it
worth analyzing. Examples of actual filtered images and of the corresponding
computed edge contours can be found in [8, 9].
The technique of Marr and Hildreth locates infinite straight edges with linear
illuminations exactly. Such edges are rather scarce in real images, and the current
paper examines how much error is introduced when the assumptions defining the
ideal case are not met.
195
0734-189X/84 $3.00
Copyright © 1984 by Academic Press, Inc.
All rights of reproduction in any form reserved.
196
VALDIS BERZINS
2. SOURCES OF ERROR
The assumptions on which Mart and Hildreth's analysis is based can break down
in the following ways.
Corners.
Edges can have finite length, ending in sharp comers. The relevant
parameters are the angle of the comer and the length of the edge.
Curves.
Edges can fail to be straight. The severity of the effect can be char-
acterized by the radius of curvature.
Nonlinearity.
The illumination can vary in a nonlinear fashion along the edge.
The relevant measure of severity is the second derivative of the illumination in the
direction of the edge.
Quantization.
Errors can be due to the finite resolution of the camera, or the
finite size of a pixel.
Noise.
The camera can record the image imperfectly, due to dust on the lens,
electrical noise, and so on.
We analyze the first three sources of error, considering each effect in isolation and
finding the dependence of the error on the parameters characterizing the error
source. The effects of the Ganssian filter are difficult to predict in the general case, so
that we have sought situations that capture the essential features of each effect and
that can be at least partially solved in closed form.
3. THE IDEAL EDGE
We briefly review the case of the infinite straight edge, introducing some of the
notations and definitions used in the rest of the analysis. Let g denote the unit
normal distribution,
1 _x2/2
g(x) = --~e
(1)
To simplify the equations, we work in coordinate systems where the unit of
measurement is equal to the scale factor of the filter, so that the standard deviation
o = 1. We can convert from a more general coordinate system
(X,Y)
with a
standard deviation o into our special coordinates using the transformation
X
X ~" m
O
Y
y.~.m
O
(2)
Let • denote the corresponding cumulative distribution
f g(t)dt,
(3)
let G denote the two dimensional unit normal distribution
O(x, y) = g(x)g(y),
(4)
ACCURACY OF LAPLACIAN EDGE DETECTORS
and let U denote the unit step function
U(x) = if x > 0 then 1 else 0.
197
(5)
An infinite sharp edge along the y axis with a linear intensity variation is
represented by the intensity function I 0 where
Io(x, y) = (ay + b)U(x)
(6)
and a and b are constants. The filtered image is given by the convolution
Fo(x,y ) = G(a, fl)Io(x - a,y - fl)dadfl.
(7)
The two dimensional integral separates into a product of one dimensional integrals.
Since • is the integral of
g, g has a
unit area, and
flg(fl)
is antisymmetric, we get
Fo(x, y) = (ay + b)'~(x).
(8)
Note that the filtering has not affected the linearity of the illumination along the
edge, which lies along the y axis. The Laplacian of F 0 is
vZF0 = -(ay
+ b)xg(x)
(9)
which is zero along x = 0, the locus of the original edge. So we see that the method
locates ideal edges exactly.
4. CORNERS
The case of a infinite right angle comer is easy to analyze. The intensity function
Ii(x, y) = U(x)U(y)
(10)
corresponds to a right angle corner in a unit step discontinuity. The corresponding
filtered image is
Fl(X , y) = Op(x)~p(y)
(11)
with a Laplacian
v2Ft -=
-(xg(x)dP(y) + yg(y)~(x)).
(12)
Note that the Laplacian is zero at the origin, indicating that the contour found by
the edge detector goes through the point of the comer exactly. Also note that the two
terms in Eq. (12) are the second directional derivatives at fight angles to the true
edges, and that they are zero along the true edge. Consequently, the displacement of
the contours is due entirely to errors introduced by approximating the directional
derivatives by the Laplacian, and not due to the Gaussian filtering. These properties
hold for infinite comers of all angles, as can be seen by examining the Laplacian in
(21) below but they do not hold for finite edges. By setting the above Laplacian to
剩余15页未读,继续阅读
资源评论
O天涯海阁O
- 粉丝: 1554
- 资源: 90
上传资源 快速赚钱
- 我的内容管理 展开
- 我的资源 快来上传第一个资源
- 我的收益 登录查看自己的收益
- 我的积分 登录查看自己的积分
- 我的C币 登录后查看C币余额
- 我的收藏
- 我的下载
- 下载帮助
最新资源
- python爬虫爬取百度百科页面
- LibreOffice Excel使用指南.pdf
- python爬虫爬取百度百科页面.zip
- Qt实现的包含图片显示功能、自适应缩放、背景图片、画roi工具 不依赖其他库纯Qt实现 在图片上可以画矩形、矩形旋转、圆形、同心圆、多边形、直线、卡尺、锚点、清空 源码: 使用Qt5.6.1-M
- 华为、华三、锐捷交换机命令生成工具
- 百度指数工具包,个人学习整理,仅供参考
- 校园闲置物品租售系统 毕业设计 源码+数据库+论文(JAVA+SpringBoot+Vue.JS).zip
- 基于Bootstrap实现的公司宣传类首页模板
- MATLAB代码:计及碳排放交易及多种需求响应的微网 电厂日前优化调度 关键词:碳排放交易 需求响应 空调负荷 电动汽车 微网 ...
- MATLAB代码:考虑P2G与碳捕集机组的多能微网低碳经济调度 关键词:碳交易 阶梯碳交易 碳捕集 多能微网 低碳调度 仿真平
- X-AnyLabeling标注工具
- 小米手机网站仿写html
- opencascade-dataset-7.7.0.tgz
- ansys联合simpack车体钢轨桥梁车轨桥耦合模型
- stylus-chrome-mv3-2.3.0-55e71c0-id.zip
- opencascade-7.5.0预编译库,使用环境window ,QT,编译器mingw32位,64位
资源上传下载、课程学习等过程中有任何疑问或建议,欢迎提出宝贵意见哦~我们会及时处理!
点击此处反馈
安全验证
文档复制为VIP权益,开通VIP直接复制
信息提交成功