Cannys
Edge Detector
Liangliang Cao
Guest Lecture
ECE 547, UIUC
Reference
John Cannys paper: A Computational
Approach to Edge Detection, IEEE Trans
PAMI, 1986
Peter Kovesis Matlab functions
http://www.csse.uwa.edu.au/~pk/Research/Matla
bFns/
Part I: Mathematical Model
Goal
Canny's aim was to discover the optimal edge
detection algorithm:
good detection the algorithm should mark as
many real edges in the image as possible.
good localization edges marked should be as
close as possible to the edge in the real image.
minimal response a given edge in the image
should only be marked once, and where possible,
image noise should not create false edges.
Formulation
Detection and Localization criterions
Impulse response of filter : f(x)
Edge : G(x)
Assuming edge is centered at x=0, filters
finite response bounded by [-W,W]
The response of the filter to the edge at its
center is given by a convolution
Criterion 1: Signal/Noise Ratio
The root mean squared noise will be measured by
Now the first criterion , the output signal-to-noise ratio
(SNR) is given by
Criterion 2: Localization
For localization we need some measure which
increases as localization increases.
Criterion 3: Multiple Response
The mean distance between adjacent maxima
in the output is twice the distance between
adjacent zero-crossings in the derivative of the
output operator.
k: the number of
noise maxima
that could lead to
a false response.
General Optimization
It is impossible to directly maximize the SNR,
localization under the multiple response
constraint.
We use penalty function: non-zero values when
one of the constraints is violated
SNR( f ) * Localizati on( f ) i Pi ( f )
where Pi is a function which has a positive value
only when a constraint is violated.
Optimization for special cases
Three kinds of edges:
Ridge edges
Roof edges (rare)
Step edges (important)
Can be get some analytical solution
for step edges?
Step Edges
Definition
Optimization for Step Edges
The problem can be solved by variational method
where filter 6 can be approximate by the first derivative
of Gaussian!
Summarization
The type of linear operator that provides the
best compromise between noise immunity and
localization is First Derivative of Gaussian
This operator corresponds to smoothing an
image with Gaussian function and then
computing the gradient
Part II: Implementation
Overview of Procedure
1) Smooth image with a Gaussian
2) Compute the Gradient magnitude using
approximations of partial derivatives
3) Thin edges by applying non-maxima
suppression to the gradient magnitude
4) Detect edges by double thresholding
Noise Reduction
First the image is convolved with a Gaussian filter
General rule for both edge and corner detection:
robust to noise
Code:
sze = ceil(6*sigma);
if ~mod(sze,2)
sze = sze+1;
end
h = fspecial('gaussian', [sze sze], sigma);
im = filter2(h, im);
Gradient
Find the first derivative in horizontal and vertical direction.
1 1 1 1
Gx Gy
1 1 1 1
(note: other more complicated filter can be used)
From this the edge gradient and direction
M [i, j ] P[i, j ]2 Q[i, j ]2
[i, j ] tan1 (Q[i, j ], P[i, j ])
The edge direction angle is rounded to one of four angles representing
vertical, horizontal and the two diagonals (0, 45, 90 and 1 degrees
for example).
Gradient (code)
A nave implementation by Matlab
[gx,gy] = gradient(im);
A better implementation by Peter
p = [0.037659 0.249153 0.426375 0.249153 0.037659];
d1 =[0.109604 0.276691 0.000000 -0.276691 -0.109604];
gx = conv2(p, d1, im, same);
gy= conv2(d1,p,im,same);
Compute gradient and direction
gradient = sqrt(gx.*gx + gy.*gy);
or = atan2(-gy, gx);
or(or<0) = or(or<0)+pi; % Map angles to 0-pi.
or = or*180/pi; % Convert to degrees.
Non-maximum suppression
Does the gradient magnitude correspond a local
maximum in the gradient direction?
The answer depends on the round angle:
[angle=0]: the edge points intensity is greater than the
intensities in the west and east directions,
[angle=90]: the edge points intensity is greater than
the intensities in the north and south directions,
[angle=135]: the edges intensity is greater than the
intensities in the north west and south east directions,
[angle=45]: the edges intensity is greater than the
intensities in the north east and south west directions.
Edge Thinning
Remove the unwanted spurious points
Use morphological operations
Code
skel = bwmorph(im,'skel',Inf);
im = im.*skel;
location = location.*skel
Example
0 0 0 0 1 1 1 3
0 0 0 1 2 1 3 1
0 0 2 1 2 1 1 0
0 1 3 2 1 1 0 0
0 3 2 1 0 0 1 0
2 3 2 0 0 1 0 1
2 3 2 0 1 0 2 1
Large gradient are potential edge points
Example
0 0 0 0 0 0 0 3
0 0 0 0 2 1 3 0
0 0 2 1 2 0 0 0
0 0 3 0 0 0 0 0 false edges
0 3 2 0 0 0 0 0
0 3 0 0 0 1 0 1
0 3 0 0 1 0 2 0
Remove those gradient that are only one pixel wide by
edge thinning
Real Example
0 90
45 135
Tracing Edges Through Hysteresis
Two thresholds Th and Tl
Keep those higher than Th
Remove those lower than Tl
For gradient between Tl and Th , only kept if they
form a continuous edge line with pixels with high
gradient magnitude
Code
aboveTl = (im > Tl);
[aboveThr, aboveThc] = find(im > Th);
% Obtain all connected regions in above T_l that include a point
% that has a value above T_h
bw = bwselect(aboveTl, aboveThc, aboveThr, 8);
Large Value of Gradient
0 0 0 0 1 1 1 3
3 0 0 1 2 1 3 1
0 0 2 1 2 1 1 0
false 0 1 3 2 1 1 0 0
edges 0 3 2 1 0 0 1 3
2 3 2 0 0 1 0 1 gaps
2 3 2 0 1 0 2 1
suppressing all values along the line of the gradient
that are not peak values
25
Parameters
1. The size of the Gaussian filter:
Smaller filters cause less blurring, and allow detection of
small, sharp lines.
A larger filter causes more blurring, smearing out the
value of a given pixel over a larger area of the image.
Larger blurring radii are more useful for detecting larger,
smoother edges (e.g., the edge of a rainbow)
2. High and low thresholds in hysteresis: allows more
flexibility than in a single-threshold approach
A threshold set too high can miss important information.
A threshold set too low will falsely identify irrelevant
information (such as noise) as important.
Real Example
Conclusion
Canny edge detector has been still arguably the
best edge detector for the last twenty years
The operator of Gradient of Gaussian has rich
theoretical meaning
Beyond edges, corner detector is more popular
in recent image recognition
SIFT talked by Mert Dikmen
