Summary of Lecture 8

• In lecture 8 we learnt how to low-pass, high-pass and band-pass filter

images in two different ways.
• We considered two filtering applications:
– Subsampling by low-pass antialiasing filters.
– Noise reduction/removal by low-pass filters

c Onur G. Guleryuz, Department of Electrical and Computer Engineering, Polytechnic University, Brooklyn, NY
° 1
 Fourier Transforms and Gibbs Phenomenon

c Onur G. Guleryuz, Department of Electrical and Computer Engineering, Polytechnic University, Brooklyn, NY
° 2
 Fourier Transforms and Gibbs Phenomenon contd.

c Onur G. Guleryuz, Department of Electrical and Computer Engineering, Polytechnic University, Brooklyn, NY
° 3
 Images and Edges

c Onur G. Guleryuz, Department of Electrical and Computer Engineering, Polytechnic University, Brooklyn, NY
° 4
 Images and Edges contd.

c Onur G. Guleryuz, Department of Electrical and Computer Engineering, Polytechnic University, Brooklyn, NY
° 5
 Edges and Noise

c Onur G. Guleryuz, Department of Electrical and Computer Engineering, Polytechnic University, Brooklyn, NY
° 6
 Edge Detection - Motivation

c Onur G. Guleryuz, Department of Electrical and Computer Engineering, Polytechnic University, Brooklyn, NY
° 7
 Edge Detection Algorithms - Continuous Time

Suppose x(t) is a continuous time signal containing “edges”. The edges can
be detected in two ways via:

1. Calculating y(t) = dx(t)

dt and finding the locations ti where |y(t)| has a
local maximum, i.e., there is an edge at location ti if:
|y(ti)| − |y(ti + ²)| ≥ 0 , ∀ 0 < |²| << 1
In order to incorporate “non-ideal” signals we will say that there is
an edge at location ti if |y(ti)| > Te where Te > 0 is a threshold or “the
strength of the edge”. However, we will still refer to this algorithm as
detecting the local maxima.

c Onur G. Guleryuz, Department of Electrical and Computer Engineering, Polytechnic University, Brooklyn, NY
° 8
 Edge Detection Algorithms - contd.

2
2. Calculating y(t) = d dtx(t)
2 and finding the locations ti of the zero crossings
of y(t), i.e., there is an edge at location ti if:
y(ti + ²) > 0 ⇒ y(ti − ²) < 0, ∀ 0 < ² << 1

or
y(ti + ²) < 0 ⇒ y(ti − ²) > 0, ∀ 0 < ² << 1

Let



 y(t) |y(t)| > Te
c(t) =  (1)
  0 otherwise
where Te > 0 is a threshold. In order to incorporate “non-ideal” sig-
nals we will say that there is an edge at location ti if c(t) has a zero
crossing at ti. We will still refer to this algorithm as detecting the
zero crossings.

c Onur G. Guleryuz, Department of Electrical and Computer Engineering, Polytechnic University, Brooklyn, NY
° 9
 Edge Detection Algorithms - Discrete Time

• We can simply approximate the derivatives d/dt and d2/dt2 by discrete

approximations. These approximations are called gradient operators.
• Images are two dimensional. If we only incorporate “row” gradient
operators we can only detect vertical edges:
– Do two passes: Detect vertical edges and then detect horizontal
edges.
– We can define “directionless” gradient operators that detect
horizontal and vertical edges in one pass.

c Onur G. Guleryuz, Department of Electrical and Computer Engineering, Polytechnic University, Brooklyn, NY
° 10
 General Algorithm

General edge detection algorithm:

1. Apply gradient operator to the image A.

2. Detect local maxima or zero crossings (with Te).
3. Generate an “edge-map” EA where



 255 pixel (m, n) is an edge pixel
EA(m, n) =  (2)
  0 otherwise
4. If a two pass approach is used to yield horizontal EhA and vertical EvA
edge maps one can define an overall map by

255 EAh (m, n) = 255


 or EAv (m, n) = 255
EA(m, n) =  (3)
 0
 otherwise

c Onur G. Guleryuz, Department of Electrical and Computer Engineering, Polytechnic University, Brooklyn, NY
° 11
 Gradient Operators

• Gradient operators are defined as filters and application of gradient

operators is implemented by linear filtering of the image.
• Note however that edge detection involves a nonlinear step where we
detect local maxima for “derivative” operators or zero crossings for
“second derivative” operators on the linearly filtered output.

c Onur G. Guleryuz, Department of Electrical and Computer Engineering, Polytechnic University, Brooklyn, NY
° 12
 Gradient Operators contd.

• Local Maxima: Determining local maxima in two dimensions is estab-

lished by comparing the absolute value of the filtered output at (m, n)
to Te.
• Zero Crossings: Let B denote the filtered and thresholded (with Te)
image.
– If B(m, n) = 0 and sign(B(m, n − 1)B(m, n + 1)) = −1 then there is a
vertical edge at pixel (m, n), etc.
– If B(m, n) 6= 0 and sign(B(m, n)B(m, n + 1)) = −1 then there is a
vertical edge between pixels (m, n) and (m, n + 1).
• Zero Crossings for Directionless Operators: On a two dimensional
grid, a zero-crossing is said to occur wherever there is a zero crossing
in at least one direction.

c Onur G. Guleryuz, Department of Electrical and Computer Engineering, Polytechnic University, Brooklyn, NY
° 13
 First Derivative Operators

   


−1 0 1  

−1 0 1 
   
• Simple: [−1 0 1], Prewitt: 


−1 0 1 , Sobel: 


−2 0 2 
   
−1 0 1 −1 0 1

Operators for horizontal edges can be obtained by transposing the filters.

c Onur G. Guleryuz, Department of Electrical and Computer Engineering, Polytechnic University, Brooklyn, NY
° 14
 Example - Vertical Edges

Detecting vertical edges on Lenna.

c Onur G. Guleryuz, Department of Electrical and Computer Engineering, Polytechnic University, Brooklyn, NY
° 15
 Example - Vertical Edges contd.

Detecting vertical edges on Lenna+10*randn(512).

c Onur G. Guleryuz, Department of Electrical and Computer Engineering, Polytechnic University, Brooklyn, NY
° 16
 Directionless Second Derivative Operators

• It can be shown that the Laplacian Operator ∇2f = ∂ 2f /∂ 2x + ∂ 2f /∂ 2y

is rotationally symmetric.
• Hence we can expect a discrete approximation to this operator to be
approximately rotationally symmetric.
 
0 1 0  

 
• One such approximation is Z(m, n) = 1 −4 1  


 
0 1 0

c Onur G. Guleryuz, Department of Electrical and Computer Engineering, Polytechnic University, Brooklyn, NY
° 17
 Edge Detection and Noise

• The derivative operators D(m, n) used in edge detection are susceptible

to noise.
• One approach would be to low pass filter (L(m, n)) an image first (to
reduce noise) and then employ the derivative operator.
• In practice one chooses spatial low pass filters in order to avoid the
Gibbs phenomenon.
• Let A(m, n) denote the original image. Then we have C = D ⊗ L ⊗ A as
the overall filtered image on which nonlinear decisions can be made.
• One can thus define a combined gradient operator H = D ⊗ L.

c Onur G. Guleryuz, Department of Electrical and Computer Engineering, Polytechnic University, Brooklyn, NY
° 18
 The Laplacian of a Gaussian Gradient Operator

2 2 2
• The Gaussian sequence defined by L(m, n) = e−(m +n )/2σ is a low-pass
sequence and can be used as a low-pass filter (σ controls the band-
width of this filter).
• The Laplacian of a Gaussian H = Z⊗L can thus be defined as a second
derivative gradient operator with favorable performance under noisy
conditions.
•σ is chosen large for very noisy conditions and small for relatively
noiseless conditions.
• Typically L(m, n) is employed as a finite extent sequence P1 × P2 (m, n
must be centered around zero −1 ≤ m, n ≤ 1 etc.). (Large values of σ
will in general require larger extents.)
• Edges are detected using the zero crossings of the filtered output (with
Te).

c Onur G. Guleryuz, Department of Electrical and Computer Engineering, Polytechnic University, Brooklyn, NY
° 19
 Example

P1 = P2 = 3,Te = 30,σ = 2

c Onur G. Guleryuz, Department of Electrical and Computer Engineering, Polytechnic University, Brooklyn, NY
° 20
 Human Visual System and Mach Bands

c Onur G. Guleryuz, Department of Electrical and Computer Engineering, Polytechnic University, Brooklyn, NY
° 21
 Low-pass Response of the Human Visual System

• The human visual system low-pass filters the scenes under observation.
• We will exploit this observation by “hiding” processing errors where
humans cannot see them.

c Onur G. Guleryuz, Department of Electrical and Computer Engineering, Polytechnic University, Brooklyn, NY
° 22
 Summary

• In this lecture we learnt about the Gibbs phenomenon.

• We learnt edge detection with:
– Gradient operators based on first derivatives and associated lo-
cal maxima.
– Gradient operators based on second derivatives and associated
zero crossings.
(please read pages 347-353 from the textbook).
• We learnt about Mach bands and the low-pass nature of the human
visual system (please read pages 51-56 from the textbook).

c Onur G. Guleryuz, Department of Electrical and Computer Engineering, Polytechnic University, Brooklyn, NY
° 23
 Homework IX

1. Detect horizontal and vertical edges in your image using the simple, Prewitt and Sobel operators.
Show horizontal edge maps, vertical edge maps and combined edge maps. Comment on differences
between horizontal and vertical edge maps. Experiment with the relevant thresholds and try to
choose them as best as you can.
2. Detect edges in your image by using the zero crossings of the Laplacian of Gaussian operator.
Adjust P1 , P2 , Te , σ to obtain the best results.
3. Repeat 2 by starting with a noisy version of your image (say with image+10randn()). Try to
adjust the parameters to match noiseless performance. Comment on the differences of the result
with 2.
4. Generate a one dimensional signal and its perceived version as I have done. Try to find a filter
which when convolved with the original signal gives the “perceived” signal. (Hint: Use DFTs of
the two signals and use the convolution → multiplication property. Be wary of divisions by zeros
or small values.) Is the filter you have discovered low-pass? Show the 1D DFT and the filter itself.

c Onur G. Guleryuz, Department of Electrical and Computer Engineering, Polytechnic University, Brooklyn, NY
° 24
References

[1] A. K. Jain, Fundamentals of Digital Image Processing. Englewood Cliffs, NJ: Prentice Hall, 1989.

c Onur G. Guleryuz, Department of Electrical and Computer Engineering, Polytechnic University, Brooklyn, NY
° 25