canny边缘检测学习笔记

Canny边缘检测的步骤：

1、高斯平滑（gaussian smoothing）

2、图像求导（derivation）

3、非极大值抑制（non-local maxima suppression）

4、滞后门限（hysteresis threshold）

* 非极大值抑制 和 滞后门限 是canny边缘检测的特色步骤

具体介绍：

http://blog.csdn.net/likezhaobin/article/details/6892176

http://www.cnblogs.com/blue-lg/archive/2011/12/25/2300899.html

算法实现（matlab源代码）：

matlab中边缘检测的函数是edge(gray_image, 'canny');我把里面关于canny算法的部分提出来写成一个单独的函数。

 

0、首先是预处理

变量a是一个m*n的矩阵，是输入的灰度图。变量e是用来保存边缘的矩阵，是一个二值图。

1 function e = canny_edge_matlab(a)
2 % Transform to a double precision intensity image if necessary
3 if ~isa(a,'double') && ~isa(a,'single') 
4   a = im2single(a);
5 end
6 [m,n] = size(a);
7 % The output edge map:
8 e = false(m,n);

变量说明

 

1、高斯平滑

GaussianDieOff是高斯衰变值，小于该值的高斯函数值都视为0.

PercentOfPixelsNotEdges代表非边缘像素占总像素个数的百分比，默认为70%。ThresholdRatio是一个系数，lowThreshhold=ThresholdRatio*highThreshold。

sigma是高斯函数里的标准差，默认为1。高斯函数的模板大小最大为60*60，实际大小由GaussianDieOff和sigma共同决定

gau保存了1D的高斯模板，用来做高斯平滑。分别沿x和y轴做平滑

dgau2D保存了2D高斯函数对x的一阶偏导，为了做图像微分。

 1 % Magic numbers
 2 GaussianDieOff = .0001;  
 3 PercentOfPixelsNotEdges = .7; % Used for selecting thresholds
 4 ThresholdRatio = .4;          % Low thresh is this fraction of the high.
 5 % Design the filters - a gaussian and its derivative
 6 sigma=1; % sigma是一个入参，默认是1，这句是我加的，为了代码可以正常运行
 7 pw = 1:30; % possible widths
 8 ssq = sigma^2;
 9 
10 % 为了确定高斯核大小，当函数值小于衰减系数时，就把他视为0
11 % 窗口大小为power(2*width)
12 width = find(exp(-(pw.*pw)/(2*ssq))>GaussianDieOff,1,'last');
13 if isempty(width)
14 width = 1;  % the user entered a really small sigma
15 end
16 t = (-width:width);
17 gau = exp(-(t.*t)/(2*ssq))/(2*pi*ssq);     % the gaussian 1D filter
18 
19 % Find the directional derivative of 2D Gaussian (along X-axis)
20 % Since the result is symmetric along X, we can get the derivative along
21 % Y-axis simply by transposing the result for X direction.
22 [x,y]=meshgrid(-width:width,-width:width);
23 dgau2D=-x.*exp(-(x.*x+y.*y)/(2*ssq))/(pi*ssq);
24 
25 % Convolve the filters with the image in each direction
26 % The canny edge detector first requires convolution with
27 % 2D gaussian, and then with the derivitave of a gaussian.
28 % Since gaussian filter is separable, for smoothing, we can us