Edge detection

 Gradient-based edge operators

 Prewitt
 Sobel
 Roberts
 Laplacian zero-crossings
 Canny edge detector
 Hough transform for detection of straight lines
 Circle Hough Transform

Digital Image Processing: Bernd Girod, © 2013 Stanford University -- Edge Detection 1
 Gradient-based edge detection
 Idea (continous-space): local gradient magnitude indicates edge strength

 Digital image:
use finite differences
to approximate
derivatives

 Edge templates

Digital Image Processing: Bernd Girod, © 2013 Stanford University -- Edge Detection 2
 Practical edge detectors
 Edges can have any orientation
 Typical edge detection scheme uses K=2 edge templates
 Some use K>2

e1  x, y 
Combination,
e.g., edge
s [x, y ] e2  x, y  ∑ ek  x, y 
2
magnitude
k
M  x, y 
or
Max k ek  x, y 

eK  x, y  α  x, y 
(edge orientation)

Digital Image Processing: Bernd Girod, © 2013 Stanford University -- Edge Detection 3
 Gradient filters (K=2)

 0 0 0   0 −1 0   [0 ] 1   [1] 0 
 −1
Central
 [0 ] 1   0
 [0 ] 0  Roberts    
Difference
 0  −1 0   0 −1 
0 0   0 1 0 

 −1 0 1   −1 −1 −1 
Prewitt  −1 [0 ] 1   0
 [0 ] 0 

 −1 0 1   1 1 1 

 −1 0 1   −1 −2 −1 
Sobel  −2 [0 ] 2   0
 [0 ] 0 

 −1 0 1   1 2 1 

Digital Image Processing: Bernd Girod, © 2013 Stanford University -- Edge Detection 4
 Kirsch operator (K=8)
 +5 +5 +5   −3 +5 +5   −3 −3 +5   −3 −3 −3 
 −3 0 −3   −3 0 +5   −3 0 +5   −3 0 +5 
Kirsch  []  []  []  [] 
 −3 −3 −3   −3 −3 −3   −3 −3 +5   −3 +5 +5 

 −3 −3 −3   −3 −3 −3   +5 −3 −3   +5 +5 −3 
 −3 0 −3   +5 0 −3   +5
 []  []  [0 ] −3   +5 [0 ] −3 
 +5 +5 +5   +5 +5 −3   +5 −3 −3   −3 −3 −3 

Digital Image Processing: Bernd Girod, © 2013 Stanford University -- Edge Detection 5
 Prewitt operator example

Magnitude of
image filtered with
 −1 0 1 
 −1
 [0 ] 1 
 −1 0 1 
(log display)

Magnitude of
image filtered with
Original
1024x710  −1 −1 −1 
 0 
 [0 ] 0 
 1 1 1 
(log display)

Digital Image Processing: Bernd Girod, © 2013 Stanford University -- Edge Detection 6
 Prewitt operator example (cont.)

Sum of squared threshold = 900 threshold = 4500 threshold = 7200

horizontal and
vertical gradients
(log display)

Digital Image Processing: Bernd Girod, © 2013 Stanford University -- Edge Detection 7
 Sobel operator example

Sum of squared threshold = 1600 threshold = 8000 threshold = 12800

horizontal and
vertical gradients
(log display)

Digital Image Processing: Bernd Girod, © 2013 Stanford University -- Edge Detection 8
 Roberts operator example

Magnitude of
image filtered with

Magnitude of
Original image filtered with
1024x710

Digital Image Processing: Bernd Girod, © 2013 Stanford University -- Edge Detection 9
 Roberts operator example (cont.)

Sum of squared threshold = 100 threshold = 500 threshold = 800

diagonal gradients
(log display)

Digital Image Processing: Bernd Girod, © 2013 Stanford University -- Edge Detection 10
 Edge orientation
Central Gradient scatter plot
Difference
 0 0 0 
 −1
[0 ] 1 

y-component

 0 0 0 
 0 −1 0 
 0
 [0 ] 0 
 0 1 0 

x-component

Roberts
 [0 ] 1 
 

y-component
 −1 0 
 [1] 0 
 
 0 −1 

x-component

Digital Image Processing: Bernd Girod, © 2013 Stanford University -- Edge Detection 11
 Edge orientation
Prewitt Gradient scatter plot
 −1 0 1 
 −1 1 
 [0 ]

y-component
 −1 0 1 
 −1 −1 −1 
 0
 [0 ] 0 
 1 1 1 

x-component

Sobel
 −1 0 1 
 −2
 [0 ] 2 

y-component
 −1 0 1 
 −1 −2 −1 
 0
 [0 ] 0 
 1 2 1 

x-component

Digital Image Processing: Bernd Girod, © 2013 Stanford University -- Edge Detection 12
 Edge orientation
Gradient scatter plot

y-component
5x5 “consistent”
gradient operator
[Ando, 2000]

x-component

Digital Image Processing: Bernd Girod, © 2013 Stanford University -- Edge Detection 13
 Gradient consistency problem

Calculate this value

using gradient field

Same result,
regardless of path?

Known
gray
value

Digital Image Processing: Bernd Girod, © 2013 Stanford University -- Edge Detection 14
 Laplacian operator
 Detect edges by considering second derivative

 Isotropic (rotationally invariant) operator

 Zero-crossings mark edge location
Detect
zero crossing

Edge profile f(x) f’(x) f”(x)

Digital Image Processing: Bernd Girod, © 2013 Stanford University -- Edge Detection 16
 Approximations of Laplacian operator by 3x3 filter
 0 1 0 
 1 1 1 
   
1  −4  1  1  −8
   1 
   
 0 1 0 
 1 1 1 

8 15

6
10
4
|H|

|H|
5
2

0 0
2 2
0 2 0 2
0 0
-2 -2 -2 -2
ω ω ωy ωx
y x

Digital Image Processing: Bernd Girod, © 2013 Stanford University -- Edge Detection 17
 Zero crossings of Laplacian

 Sensitive to very fine detail and noise  blur image first

 Responds equally to strong and weak edges
 suppress zero-crossings with low gradient magnitude

Digital Image Processing: Bernd Girod, © 2013 Stanford University -- Edge Detection 18
 Laplacian of Gaussian
 Filtering of image with Gaussian and Laplacian operators can be combined into
convolution with Laplacian of Gaussian (LoG) operator
σ= 2
0.02
x2 + y2
1  x +y  − 2 2
LoG (x, y )= − 4 
1− 2 
e 2σ 2 0

πσ  2σ  -0.02

-0.04

-0.06

-0.08

-0.1
4
2 4
0 2
0
-2
-2
Y -4 -4 X

Digital Image Processing: Bernd Girod, © 2013 Stanford University -- Edge Detection 19
 Discrete approximation of Laplacian of Gaussian
σ= 2
0 0 1 2 2 2 1 0 0

0 2 3 5 5 5 3 2 0 10 200
1 3 5 3 0 3 5 3 1
0 150
2 5 3 -12 -23 -12 3 5 2
-10
100

|H|
2 5 0 -23 -40 -23 0 5 2
-20
2 5 3 -12 -23 -12 3 5 2
50
1 3 5 3 0 3 5 3 1 -30

0 2 3 5 5 5 3 2 0 -40 0
4
0 0 1 2 2 2 1 0 0 2 4 2
0 2 0 2
0 0
-2 -2
-2 -2
Y -4 -4 X ωy ωx

Digital Image Processing: Bernd Girod, © 2013 Stanford University -- Edge Detection 20
 Zero crossings of LoG

σ= 2 σ =2 2 σ =4 2 σ =8 2

Digital Image Processing: Bernd Girod, © 2013 Stanford University -- Edge Detection 21
 Zero crossings of LoG – gradient-based threshold

σ= 2 σ =2 2 σ =4 2 σ =8 2

Digital Image Processing: Bernd Girod, © 2013 Stanford University -- Edge Detection 22
 Canny edge detector
1. Smooth image with a Gaussian filter
2. Approximate gradient magnitude and angle (use Sobel, Prewitt . . .)

∂f ∂f 
2 2
∂f  ∂f 
M  x, y  ≈   +   α  x, y  ≈ tan 
−1

 ∂x   ∂ y   ∂y ∂x 

3. Apply nonmaxima suppression to gradient magnitude

4. Double thresholding to detect strong and weak edge pixels
5. Reject weak edge pixels not connected with strong edge pixels

[Canny, IEEE Trans. PAMI, 1986]

Digital Image Processing: Bernd Girod, © 2013 Stanford University -- Edge Detection 23
 Canny nonmaxima suppression
 Quantize edge normal to one of four directions:
horizontal, -45o, vertical, +45o
 If M[x,y] is smaller than either of its neighbors in edge normal direction
 suppress; else keep.

[Canny, IEEE Trans. PAMI, 1986]

Digital Image Processing: Bernd Girod, © 2013 Stanford University -- Edge Detection 24
 Canny thresholding and suppression of weak edges
 Double-thresholding of gradient magnitude

 Typical setting:
 Region labeling of edge pixels
 Reject regions without strong edge pixels

[Canny, IEEE Trans. PAMI, 1986]

Digital Image Processing: Bernd Girod, © 2013 Stanford University -- Edge Detection 25
 Canny edge detector

σ= 2 σ =2 2 σ =4 2

Digital Image Processing: Bernd Girod, © 2013 Stanford University -- Edge Detection 26
 Hough transform
 Problem: fit a straight line (or curve) to a set of edge pixels
 Hough transform (1962): generalized template matching technique
 Consider detection of straight lines y = mx + c

x m

edge pixel

y
c
Digital Image Processing: Bernd Girod, © 2013 Stanford University -- Edge Detection 28
 Hough transform (cont.)
 Subdivide (m,c) plane into discrete “bins,” initialize all bin counts by 0
 Draw a line in the parameter space [m,c] for each edge pixel [x,y] and increment
bin counts along line.
 Detect peak(s) in [m,c] plane
x m

detect peak

y
c
Digital Image Processing: Bernd Girod, © 2013 Stanford University -- Edge Detection 29
 Hough transform (cont.)
 Alternative parameterization avoids infinite-slope problem
x ρ
θ
edge pixel
ρ

−−ππ π
x cosθ + y sin θ = ρ 2 2 θ
y

 Similar to Radon transform

Digital Image Processing: Bernd Girod, © 2013 Stanford University -- Edge Detection 30
 Hough transform example

100

200

300

400

500

600

700

800
200 400 600 800

Original image

Digital Image Processing: Bernd Girod, © 2013 Stanford University -- Edge Detection 31
 Hough transform example

100

200

300

400

500

600

700

800
200 400 600 800

Original image

Digital Image Processing: Bernd Girod, © 2013 Stanford University -- Edge Detection 32
 Hough transform example

100

200

300

400

500

600

700

800
200 400 600 800

Original image

Digital Image Processing: Bernd Girod, © 2013 Stanford University -- Edge Detection 33
 Hough transform example
-4000 1500

-2000
ρ 1000
0

500
2000
De-skewed
Paper Paper
(3264x2448)

4000 0
-50 0 50
θ
-81.8 deg

-81.8 deg

Global thresholding

Digital Image Processing: Bernd Girod, © 2013 Stanford University -- Edge Detection 34
 Circle Hough Transform
 Find circles of fixed radius r
x x0
Circle center

edge pixel

y
y0
 Equivalent to convolution (template matching) with a circle

Digital Image Processing: Bernd Girod, © 2013 Stanford University -- Edge Detection 35
 Circle Hough Transform for unknown radius
 3-d Hough transform for parameters (x0 , y0 ,r )
 2-d Hough transform aided by edge orientation  “spokes” in parameter space

x x0

Circle center

edge pixel

y
y0

Digital Image Processing: Bernd Girod, © 2013 Stanford University -- Edge Detection 36
 Example: circle detection by Hough transform
Prewitt edge detection Detected circles
Original
coins image

Digital Image Processing: Bernd Girod, © 2013 Stanford University -- Edge Detection 37