Image Processing 1, Exercise 3

1 The Dirac Delta

[basic] This exercise is an exploration of convolutions with the 2D Dirac delta to help you understand
and build your intuition about 2D convolutions. As usual, you can rely on the tables in the back of
your course notes.

Let x be a vector in R2 .
  
2
(a) Let f (x) = δ x − . Sketch rect ∗ f .
1

Solution: Use the convolution property of the Dirac delta! Let us denote δ(x − x0 ) = δx0 .
Then the convolution property tells us: (g ∗ δx0 )(x) = g(x − x0 ).
 2 
Thus we have (rect ∗ f )(x) = rect x − . Hence the support of rect ∗ f is a unit square
1
centered at (2, 1).
2
x2

1 1

0
0 1 2 3
x1
  
−3
(b) Let f (x) = δ −x − . Use the Fourier transform to show that δ(x) = δ(−x) and use
1
this fact to sketch rect ∗ f .

Solution:
Image space Fourier space
f (x) fˆ(ω)
f (−x) fˆ(−ω)
δ(x) 1
δ(−x) 1
     
−3 −3
Therefore, we deduce that δ(x) = δ(−x). Hence, δ −x − = δ x+ =
1 1
  
3
δ x− . Then use the same strategy as in (a).
−1
0

−1
x2

−2
1 2 3 4
x1

Page 1 of 4
 (c) Fix x0 ∈ R2 . Prove that f ∗ δ(· − x0 ) = f (· − x0 ) using the Fourier transform.

F
Solution: By the convolution theorem, we have f ∗ δ(· − x0 ) ←→ fˆ(ω)e−j⟨ω,x0 ⟩ . Taking
the inverse Fourier transform and using the translation/modulation property of the Fourier
transform (Chapter 1, Slide 26) we find f ∗ δ(· − x0 ) = f (· − x0 ).
   
2 −3
(d) Let x1 = and x2 = . Use the result of (c) to compute g = δ(· − x1 ) ∗ δ(x2 + ·). Then,
1 1
plot g ∗ rect.

·) = δ(· + x2 ) = δ(· − (−x2 )). Therefore, using the result of

Solution: Note thatδ(x2+
5
(c) we find g(x) = δ x − . The plot of g ∗ rect is below.
0
1
x2

0 1

−1
3 4 5 6
x1

(e) Let f ∈ L1 (Rd ), which ensures that the Fourier transform fˆ is a continuous function. In this
part you will prove the identity

(2π)d fˆ(ω)δ(ω − ω0 ) = (2π)d fˆ(ω0 )δ(ω − ω0 ), (1)

which is the dual form of the Multiplication identity on Slide 23 in Chapter 1. We will proceed
in two steps:
(i) First prove that f ∗ e j⟨ω0 ,·⟩ = fˆ(ω0 )e j⟨ω0 ,·⟩ by an explicit calculation.
(ii) Prove the identity (1) using the result of (i). Hint: Take the inverse Fourier transform of
the left-hand side of (1).

Solution: (i) Since f ∈ L1 (Rd ), we see that

Z Z
(f ∗ e j⟨ω0 ,·⟩ )(x) = f (y)e j⟨ω0 ,x−y⟩ dy = e j⟨ω0 ,x⟩ f (y)e−j⟨ω0 ,y⟩ dy = e j⟨ω0 ,x⟩ fˆ(ω0 ).
Rd Rd

(ii) Taking the inverse Fourier transform of the left-hand side is f ∗ e j⟨ω0 ,·⟩ (convolution
theorem). By (i) it equals fˆ(ω0 )e j⟨ω0 ,·⟩ . Taking the Fourier transform of this yields the
right-hand side.

2 Aliasing
[intermediate] Aliasing must be considered in all practical systems that involve sampling continuous
signals, e.g., cameras and scanning microscopes. In this exercise, we illustrate the effect of aliasing
in the space domain and analyse the effect in the Fourier domain.

The image f (x) = cos 2π 32 x1 + cos 2π 53 x2 , with x ∈ R2 , is sampled at x = kT with k ∈ Z2 .

The results for various values of T are shown in Figure 1.

Remark: the origin of the coordinate system is at the top left corner of the image, x1 -axis goes
horizontally to the right hand side and x2 -axis goes vertically to the bottom of the page.
(a) Match each of the images in Figure 1 with the correct value of T from the list T = 1/6, 1, 2, 3,
and 6.

Page 2 of 4
 (i) (ii) (iii) (iv) (v)

Figure 1: Sampled versions of f (x).

Solution: Answer: T = 6, 1/6, 2, 1, and 3.

How to solve: Start with (i), (iii), and (v), since these are the ones that are constant in
either the x1 or x2 direction. For example, only T = 6 makes both cos(2π 23 k1 T ) and
cos(2π 35 k2 T ) constant, so (i) is 6. Similar logic works for (iii) and (v). Then, note that (iv)
has a period of 2 in the x1 direction, only one of the remaining T values gives this. Image
(ii) follows by process of elimination.

(b) Determine the Fourier transforms of g(x) = cos(2π 23 x) and h(x) = cos(2π 53 x), then plot them
in the range [−4π, 4π].

Solution: From the 1D Fourier transform table: ĝ(ω) = π(δ(ω + 3π) + δ(ω − 3π)) and
ĥ(ω) = π(δ(ω + 10 10
3 π) + δ(ω − 3 π)).
ĝ(ω)

π π

−3π 0 3π
ω
ĥ(ω)

π π

− 10 0 10
3 π 3 π
ω
(Note that Fourier transforms are usually complex-valued, so we would have to either plot
the real and imaginary parts separately, or plot the magnitude and phase separately. But,
for these functions, the Fourier transform happened to be real.)

(c) Let g1 (x) and h1 (x) be the signals formed by multiplying functions g and h from (b) respectively
with a Dirac comb (with sampling step T = 1). Determine the Fourier transforms of g1 and h1 ,
plot them in the range [−4π, 4π].
P ∗ P
Solution: ĝ1 (ω) = n∈Z π(δ(ω − 2πn + 3π) + δ(ω − 2πn − 3π)) = n∈Z 2πδ(ω − 2πn + π)
and ĥ1 (ω) = n∈Z π(δ(ω − 2πn + 34 π) + δ(ω − 2πn − 34 π)).
P

Explanation for the * step:

X X
δ(ω − 2πn + 3π) = δ(ω − 2π(n + 1) + 3π)
n∈Z n∈Z
X
= δ(ω − 2πn + π).
n∈Z

Page 3 of 4
 and
X X
δ(ω − 2πn − 3π) = δ(ω − 2π(n − 2) − 3π)
n∈Z n∈Z
X
= δ(ω − 2πn + π).
n∈Z

If your answers look different from these, remember that you can freely make variable
substitutions in the sums like above, so your answers might be completely correct. The
best way to check your answer is to plot it and compare with the solution plots.
ĝ1 (ω)

2π 2π 2π 2π

−3π −π 0 π 3π
ω
ĥ1 (ω)

π π π π π π π π

− 10 8
− 43 π − 23 π 0 2 4 8 10
3 π −3π 3π 3π 3π 3 π
ω

(d) Based on your previous answers, which (if either) of g and h can be reconstructed exactly from
its samples with T = 1?

Solution: Neither, since both exhibit aliasing (that is, the central part of ĝ1 (ω) doesn’t
match that of ĝ(ω), same for ĥ1 (ω)).

(e) What is the range of T such that both g and h can be perfectly reconstructed from their samples?

Solution: According to the Nyquist criterion, ωmax < Tπ is required to ensure alias-free.
ωg = 3π, ωh = 10 10 π 3 3
3 π so ωmax = 3 π and T < ωmax = 10 . Hence the range of T is (0, 10 ).

Page 4 of 4