---
title: "边缘检测算法原理"
author: "mywebcode"
date: "2013-11-05T14:13:00Z"
tags: []
slug: "untitled"
draft: false
lastmod: "2014-04-30T09:25:27Z"
id: 2058468
url: "https://mywebcode.iteye.com/blog/2058468"
---

边缘检测算法有如下四个步骤：

滤波：边缘检测算法主要是基于图像强度的一阶和二阶导数，但导数的计算对噪声很敏感，因此必须使用滤波器来改善与噪声有关的边缘检测器的性能．需要指出，大多数滤波器在降低噪声的同时也导致了边缘强度的损失，因此，增强边缘和降低噪声之间需要折衷．

增强：增强边缘的基础是确定图像各点邻域强度的变化值．增强算法可以将邻域（或局部）强度值有显著变化的点突显出来．边缘增强一般是通过计算梯度幅值来完成的．

检测：在图像中有许多点的梯度幅值比较大，而这些点在特定的应用领域中并不都是边缘，所以应该用某种方法来确定哪些点是边缘点．最简单的边缘检测判据是梯度幅值阈值判据．

定位：如果某一应用场合要求确定边缘位置，则边缘的位置可在子像素分辨率上来估计，边缘的方位也可以被估计出来．

在边缘检测算法中，前三个步骤用得十分普遍。这是因为大多数场合下，仅仅需要边缘检测器指出边缘出现在图像某一像素点的附近，而没有必要指出边缘的精确位置或方向．边缘检测误差通常是指边缘误分类误差，即把假边缘判别成边缘而保留，而把真边缘判别成假边缘而去掉．边缘估计误差是用概率统计模型来描述边缘的位置和方向误差的．我们将边缘检测误差和边缘估计误差区分开，是因为它们的计算方法完全不同，其误差模型也完全不同．  
最近的二十年里发展了许多边缘检测器，这里仅讨论常用的几种边缘检测器．  
1、 Roberts算子Roberts交叉算子为梯度幅值计算提供了一种简单的近似方法：

![](http://img.blog.csdn.net/20131107135234640?watermark%2F2%2Ftext%2FaHR0cDovL2Jsb2cuY3Nkbi5uZXQvZXNyaWNoaW5hY2Q=%2Ffont%2F5a6L5L2T%2Ffontsize%2F400%2Ffill%2FI0JBQkFCMA%3D%3D%2Fdissolve%2F70%2Fgravity%2FSouthEast)(1)

用卷积模板，上式变成：

![](http://img.blog.csdn.net/20131107135250828?watermark%2F2%2Ftext%2FaHR0cDovL2Jsb2cuY3Nkbi5uZXQvZXNyaWNoaW5hY2Q=%2Ffont%2F5a6L5L2T%2Ffontsize%2F400%2Ffill%2FI0JBQkFCMA%3D%3D%2Fdissolve%2F70%2Fgravity%2FSouthEast)(2)

其中Gx和Gy由下面的模板计算：

![](http://img.blog.csdn.net/20131107135330484?watermark%2F2%2Ftext%2FaHR0cDovL2Jsb2cuY3Nkbi5uZXQvZXNyaWNoaW5hY2Q=%2Ffont%2F5a6L5L2T%2Ffontsize%2F400%2Ffill%2FI0JBQkFCMA%3D%3D%2Fdissolve%2F70%2Fgravity%2FSouthEast)![](http://img.blog.csdn.net/20131107135330484?watermark%2F2%2Ftext%2FaHR0cDovL2Jsb2cuY3Nkbi5uZXQvZXNyaWNoaW5hY2Q=%2Ffont%2F5a6L5L2T%2Ffontsize%2F400%2Ffill%2FI0JBQkFCMA%3D%3D%2Fdissolve%2F70%2Fgravity%2FSouthEast)(3)

同前面的梯度算子一样，差分值将在内插点处计算．Roberts算子是该点连续梯度的近似值，而不是所预期的点处的近似值．Roberts边缘检测器的试验结果见图（3） .  
2、 Sobel算子正如前面所讲，采用3\*3邻域可以避免在像素之间内插点上计算梯度．考虑一下图6．2中所示的点周围点的排列．Sobel算子也是一种梯度幅值，

![](http://img.blog.csdn.net/20131107135403593?watermark%2F2%2Ftext%2FaHR0cDovL2Jsb2cuY3Nkbi5uZXQvZXNyaWNoaW5hY2Q=%2Ffont%2F5a6L5L2T%2Ffontsize%2F400%2Ffill%2FI0JBQkFCMA%3D%3D%2Fdissolve%2F70%2Fgravity%2FSouthEast)(4)

其中的偏导数用下式计算：

![](http://img.blog.csdn.net/20131107135423343?watermark%2F2%2Ftext%2FaHR0cDovL2Jsb2cuY3Nkbi5uZXQvZXNyaWNoaW5hY2Q=%2Ffont%2F5a6L5L2T%2Ffontsize%2F400%2Ffill%2FI0JBQkFCMA%3D%3D%2Fdissolve%2F70%2Fgravity%2FSouthEast)(5)

其中常数C=2;  
![](http://img.blog.csdn.net/20131107135503515?watermark%2F2%2Ftext%2FaHR0cDovL2Jsb2cuY3Nkbi5uZXQvZXNyaWNoaW5hY2Q=%2Ffont%2F5a6L5L2T%2Ffontsize%2F400%2Ffill%2FI0JBQkFCMA%3D%3D%2Fdissolve%2F70%2Fgravity%2FSouthEast)![](http://img.blog.csdn.net/20131107135512437?watermark%2F2%2Ftext%2FaHR0cDovL2Jsb2cuY3Nkbi5uZXQvZXNyaWNoaW5hY2Q=%2Ffont%2F5a6L5L2T%2Ffontsize%2F400%2Ffill%2FI0JBQkFCMA%3D%3D%2Fdissolve%2F70%2Fgravity%2FSouthEast)和其他的梯度算子一样，和可用卷积模板来实现：

![](http://img.blog.csdn.net/20131107135536406?watermark%2F2%2Ftext%2FaHR0cDovL2Jsb2cuY3Nkbi5uZXQvZXNyaWNoaW5hY2Q=%2Ffont%2F5a6L5L2T%2Ffontsize%2F400%2Ffill%2FI0JBQkFCMA%3D%3D%2Fdissolve%2F70%2Fgravity%2FSouthEast)![](http://img.blog.csdn.net/20131107135544734?watermark%2F2%2Ftext%2FaHR0cDovL2Jsb2cuY3Nkbi5uZXQvZXNyaWNoaW5hY2Q=%2Ffont%2F5a6L5L2T%2Ffontsize%2F400%2Ffill%2FI0JBQkFCMA%3D%3D%2Fdissolve%2F70%2Fgravity%2FSouthEast)(6)

请注意这一算子把重点放在接近于模板中心的像素点．图6.3和图6.4表明了这一算子的作用．Sobel算子是边缘检测器中最常用的算子之一．  
![](http://img.blog.csdn.net/20131107135620125?watermark%2F2%2Ftext%2FaHR0cDovL2Jsb2cuY3Nkbi5uZXQvZXNyaWNoaW5hY2Q=%2Ffont%2F5a6L5L2T%2Ffontsize%2F400%2Ffill%2FI0JBQkFCMA%3D%3D%2Fdissolve%2F70%2Fgravity%2FSouthEast)![](http://img.blog.csdn.net/20131107135631906?watermark%2F2%2Ftext%2FaHR0cDovL2Jsb2cuY3Nkbi5uZXQvZXNyaWNoaW5hY2Q=%2Ffont%2F5a6L5L2T%2Ffontsize%2F400%2Ffill%2FI0JBQkFCMA%3D%3D%2Fdissolve%2F70%2Fgravity%2FSouthEast)

上图 用于说明Sobel算子和Prewitt算子的邻域像素点标记

3、 Prewitt算子 ,Prewitt算子与Sobel算子的方程完全一样，只是常量c=1．所以

![](http://img.blog.csdn.net/20131107135707890?watermark%2F2%2Ftext%2FaHR0cDovL2Jsb2cuY3Nkbi5uZXQvZXNyaWNoaW5hY2Q=%2Ffont%2F5a6L5L2T%2Ffontsize%2F400%2Ffill%2FI0JBQkFCMA%3D%3D%2Fdissolve%2F70%2Fgravity%2FSouthEast)![](http://img.blog.csdn.net/20131107135719359?watermark%2F2%2Ftext%2FaHR0cDovL2Jsb2cuY3Nkbi5uZXQvZXNyaWNoaW5hY2Q=%2Ffont%2F5a6L5L2T%2Ffontsize%2F400%2Ffill%2FI0JBQkFCMA%3D%3D%2Fdissolve%2F70%2Fgravity%2FSouthEast)(7)

　　请注意，与Sobel算子不同，这一算子没有把重点放在接近模板中心的像素点．这种边缘检测器的性能也示意在图6.3和图6.4中．

**4、 各种算法比较**现在来比较一下上面讨论过的边缘检测器．我们将按照本节开头所提出的滤波、增强和检测这三个步骤，来比较各种方法．第四步定位将不讨论．首先给出在忽略滤波步骤情况下Roberts、Sobel和Prewitt边缘检测方法实验结果，如图6.3所示．对滤波后的图象进行边缘检测的结果见图6.4，其中滤波器为前一章介绍的![](http://b.hiphotos.baidu.com/space/pic/item/0df431adcbef7609c22447312edda3cc7cd99e75.jpg)高斯滤波器，梯度幅值的计算见方程6.3．比较图6.3和图6.4可以发现，由于噪声影响，一些假边缘也被检测出来了．

![](http://img.blog.csdn.net/20131107135143687?watermark%2F2%2Ftext%2FaHR0cDovL2Jsb2cuY3Nkbi5uZXQvZXNyaWNoaW5hY2Q=%2Ffont%2F5a6L5L2T%2Ffontsize%2F400%2Ffill%2FI0JBQkFCMA%3D%3D%2Fdissolve%2F70%2Fgravity%2FSouthEast)

图 用于边缘检测的测试图像

(a)原始图像(b)7x7高斯滤波的图像

![](http://img.blog.csdn.net/20131107135154406?watermark%2F2%2Ftext%2FaHR0cDovL2Jsb2cuY3Nkbi5uZXQvZXNyaWNoaW5hY2Q=%2Ffont%2F5a6L5L2T%2Ffontsize%2F400%2Ffill%2FI0JBQkFCMA%3D%3D%2Fdissolve%2F70%2Fgravity%2FSouthEast)

图 各种边缘检测器对未经滤波的图像（a）进行边缘检测的比较．

(c)Roberts交叉算子．（d）Sobel算子．(e)Prewitt算子．

![](http://img.blog.csdn.net/20131107135207875?watermark%2F2%2Ftext%2FaHR0cDovL2Jsb2cuY3Nkbi5uZXQvZXNyaWNoaW5hY2Q=%2Ffont%2F5a6L5L2T%2Ffontsize%2F400%2Ffill%2FI0JBQkFCMA%3D%3D%2Fdissolve%2F70%2Fgravity%2FSouthEast)

图 各种边缘检测器对滤波后的图像（图6.3(b)）进行边缘检测的结果．

(a)Roberts交叉算子．（b）Sobel算子．(c)Prewitt算子