• Linear shift invariant (LSI) system

– Convolution
• Fourier transform(FT) and its properties
– Definition, shifting, scaling
– FT and convolution
– Filtering with FT
Object and image
1.2

0.8

0.6

0.4

0.2

-0.2
0 10 20 30 40 50 60 70
Object and image
1

0.9

0.8

0.7

0.6

0.5

0.4

0.3

0.2

0.1

0
20 25 30 35 40 45
 Systems
• A system is a transformation of the input
signal f(x,y) to an output signal g(x,y).
• g(x,y) = S[f(x,y)]
• For example, in CT, f(x,y)=µ(x,y), the
attenuation coefficient of tissues, while g(x,y)
is the reconstructed CT images. It can be
divided into two subsystems.
 Linear System
A system is a linear system if, when the input consists of a
weighted summation of several signals, the output will also be a
weighted summation of the responses of the system to each
individual input signal.
 Impulse response
When there is a point object δ(x-x’,y-y’) at (x’,y’) as the input to a
system, the output h(x, y;x’,y’) is called impulse response
function or point spread function.

Superposition integral.
 A point object and image
Point Spread Function (PSF)/impulse response function

Point Spread Function (PSF) : the image obtained by a system for a point object
located at the origin.
Point Objects and image
 Linear–shift invariant (LSI) system
A system is shift-invariant if an arbitrary translation of the input
results in an identical translation in the output .

For a LSI system, h(x,y;x’,y’) can be reduced to

where h(x,y) is the psf when the object is located at (0,0).

 Linear–shift invariant (LSI) system
The image g (x,y) of an arbitrary object f(x,y ) is the
convolution of the object f(x,y) and the point spread function
h(x,y).
1D convolution is defined as
2D convolution is defined as

Input image/ object to the system Output image of the system:

Imaging
System
 Two methods to do discrete
convolution
1. Flip one object and drag it across the other.

2. Replace each point in f1(n) with the other

object f2 weighted by f1(n).
Clicker question
(Method 1) What is the largest element in the convolution of
[2 3 3] with
[ 1 2]?

(a) 5
(b) 8
(c) 6
(d)9
Clicker question
(Method 2) What is the largest element in the convolution of
[2 3 3] with
[ 1 2]?

(a) 5
(b) 8
(c) 6
(d)9
Clicker question
What is the third element in the convolution of [2 3 3 1 5 2]
with [ 1 2 -1 5]?

(a) 5
(b) 8
(c) 7
(d)9
 One-dimensional convolution example:
g(x) = rect(x)*rect (x/2)
Recall: rect (x) = 1 for |x| < ½

rect (x/2) = 1 for |x|/2 < ½

or |x| < 1

 x 
g ( x )   rect    rect( x  x) dx Flip one object and
 2 drag across the other.


flip delay
 x   
g ( x)   rect    rect(( x  x))dx
 2
Convolution by integration
https://www.youtube.com/watch?v=MEDjw6VcDTY
One-dimensional convolution example, continued:
Case 1: no overlap of rect(x-x’) and rect (x’/2)

3
g ( x)  0 for x-
2

Case 2: partial overlap of rect (x-x’) and rect (x’/2)

x1/ 2
g(x)  11 dx
1


x1/ 2 1
 x 1  x   1
2
3 3 1
g(x)  x  for -  x  -
2 2 2
One-dimensional convolution example, continued(2):
x1/ 2
Case 3: complete overlap
g(x)  
x
11 dx
1/ 2
1
g(x)  1 for - 2 x 2 1

1
Case 4: partial overlap g(x)  
x
11 dx
1/ 2

 x 1x1/ 2 1 (x  12 )
1
g(x)  x  3
2
for 2 x 2 3

Case 5: no overlap
g ( x)  0 for x  23
One-dimensional convolution example, continued(3):
Result of convolution:

3
g ( x)  0 for x-
2
3 3 1
g ( x)  x  for -  x  -
2 2 2
1 1
g ( x)  1 for -  x 
2 2
3 1 3
g ( x)   x  for x
2 2 2
3
g ( x)  0 for x
2
Use Convolution to describe LSI system
The image g (x,y) of an arbitrary object f(x,y ) is the
convolution of the object f(x,y) and the point spread function
h(x,y).

Input image/ object to the system Output image of the system:

Imaging
System
 1D (one dimension) Fourier transform
1D (one dimension) Fourier transform: A function f(x) is the
superposition of many sinusoidal waves at different spatial
frequencies u with the corresponding weight of F(u).

F(u) is the Fourier transform of f(x), and f(x) is the inverse

Fourier transform of F(u).

f ( x)   F (u)e j 2 ux du


F (u )   f ( x)e j 2 ux dx

 Magnitude and phase of the spectrum
magnitude of spectrum
6 2
magnitude+phase
5
Phase only
Magnitude only
1.5
4

3 1

0.5
1

0
0 5 10 15 20 25 30 35 40 0

Phase of spectrum
3

-0.5
0 5 10 15 20 25 30 35 40
2

F(u) is usually a complex function. So we need to

1
display the magnitude and the phase separately.
0

-1

-2
https://www.youtube.com/watch?v=r18Gi8l
-3
0 5 10 15 20 25 30 35 40
SkfM An excellent animation on the phase
of FT.
 2D FT
2D (two dimension) Fourier transform:
An image f(x,y) is the superposition of
many images of sinusoidal waves at
different spatial frequencies (u,v) with
the corresponding weight of F(u,v).
F(u,v) is the Fourier transform or spec
trum of f(x,y), and f(x,y) is the inverse
Fourier transform of F(u,v).
(x,y) is called image space.
(u,v) is called k-space, frequency space.

Cos[2π (uox+voy)]
f ( x, y)   F (u, v)e j 2 ( ux  vy )
dudv

F (u, v)   f ( x, y)e j 2 (ux vy ) dxdy

 Fourier Transform of delta function
What is the fourier transform of δ(x)?

A Fourier transform pair:

Clicker question
What is the fourier transform of δ(x-1)?

(a) exp(2πui)
(b) exp(-2πui)
(c) 1
(d)0
Fourier Transform of complex
exponential
Fourier Transform of sine and cos
Find the FT of rect(x).
Clicker question
What is the fourier transform of 2cos(6x)?

(a) δ(u-3/π)+ δ(u+3/π)

(b) δ(u-6/π)+ δ(u+6/π)
(c) δ(u-3)+ δ(u+3)
(d)δ(u-6)+ δ(u+6)
Basic Fourier transform pairs (Table 2.1 in Prince’s Textbook)
Properties
of Fourier
Transform
(Table 2.2
in Prince’s
Textbook) Real component
Imaginary component
Magnitude
Phase
 Argument shifting in FT
Given that the FT of f(x) is F(u), find FT of f(x-x0) in terms of F(u).
Clicker question
What is the fourier transform of exp(-2x2)?
 Argument scaling in FT
Given that the FT of f(x) is F(u), find FT of f(x/a) in terms of F(u).

The more concentrated f(x) is, the more spread out F(u) is.
In particular, if we "squeeze" a function in x, it spreads out in u and
vice-versa.
 FT of a real function

two-Sided Amplitude Spectrum of y(t)

two-Sided Amplitude Spectrum of y(t)
1
1

0.9
0.9

0.8
0.8

0.7
0.7

0.6 0.6
|Y(f)|
0.5
|Y(f)|

0.5

0.4 0.4

0.3 0.3

0.2 0.2

0.1 0.1

0 0
0 100 200 300 400 500 600 700 800 900 1000 -500 -400 -300 -200 -100 0 100 200 300 400 500
Frequency (Hz) Frequency (Hz)

Amplitude of the original spectrum Amplitude of the spectrum after using fftshift
 Convolution by Fourier transform
• Fourier transform of the convolution of two
signals (images) equals the product of the
individual Fourier transforms.
• Given the FT of f1(x) and f2(x) are F1(u) and
F2(u), respectively, and f3(x)=f1(x)*f2(x),
where * means convolution,
F3(u)= F1(u) F2(u)
 Convolution by Fourier transform
• Given f1(x) and f2(x), find f3=f1*f2

• A. Find F1(u) and F2(u)

• B. F3(u)= F1(u) X F2(u)
• C. Find f3(x) by inverse Fourier transform
 Parseval’s theorem
• The energy in the time/space domain equals
that in the frequency domain.
 Reading material
• The Fourier Transform and its Applications
see.stanford.edu/materials/lsoftaee261/book-
fall-07.pdf

• Chapter 2 of A. C. Kak and Malcolm Slaney,

Principles of Computerized Tomographic
Imaging, Society of Industrial and Applied
Mathematics, 2001.(Available on-line)