Chapter 04a Frequency Filtering (Fundamentals)

Digital Image Processing
Filtering in the Frequency Domain

(Fundamentals)
Christophoros Nikou
cnikou@cs.uoi.gr
University of Ioannina - Department of Computer Science

2 Filtering in the Frequency Domain
Filter: A device or material for suppressing or
minimizing waves or oscillations of certain
frequencies.
Frequency: The number of times that a periodic

function repeats the same sequence of values
during a unit variation of the independent variable.
Webster’s New Collegiate Dictionary
C. Nikou – Digital Image Processing (E12)

3 Jean Baptiste Joseph Fourier
Fourier was born in Auxerre,
France in 1768.
– Most famous for his work “La
Théorie Analitique de la
Chaleur” published in 1822.
– Translated into English in 1878:
“The Analytic Theory of Heat”.
Nobody paid much attention when the work
was first published.
One of the most important mathematical
theories in modern engineering.
4 The Big Idea
Images taken from Gonzalez & Woods, Digital Image Processing (2002)
Any function that periodically repeats itself can

be expressed as a sum of sines and cosines of
different frequencies each multiplied by a
different coefficient – a Fourier series
5 1D continuous signals
• It may be
considered both as
continuous and
discrete.
• Useful for the
representation of , x  x0
discrete signals  ( x  x0 )  
 0 otherwise
through sampling

of continuous
signals.


f ( x) ( x  x0 )dx  f ( x0 )

6 1D continuous signals (cont.)
Impulse train function


ST (t )    (t  nT )
n 
 
x[n]  x(t ) S T (t )   x(t ) (t  nT )   x(nT ) (t  nT )
n  n 

x[n]  x(t ) ST (t )

  x(t ) (t  nT )
n 

  x(nT ) (t  nT )
n 

• The Fourier series expansion of a periodic

signal f (t).
 2
ce
j nt
f (t )  n
T
n 
T /2 2
1 j nt
cn  
T T / 2
f (t )e T
dt

• The Fourier transform of a continuous

signal f (t).

F ( )  
 j 2 t
f (t )e dt


 F (  )e d
j 2t
f (t ) 

• Attention: the variable is the frequency (Hz) and
not the radial frequency (Ω=2πμ) as in the Signals
and Systems course.
sin(W )
f (t )  PW / 2 (t )  F (  )  W
(W )
• Convolution property of the FT.


f (t )* h(t )  

f ( )h(t   )d
f (t )* h(t )  F (  ) H (  )
f (t )h(t )  F (  ) * H (  )

• Intermediate result
− The Fourier transform of the impulse train.

 
1  n 

n 
 (t  nT ) 
T
   
n  

T 
• It is also an impulse train in the frequency

domain.
• Impulses are equally spaced every 1/ΔΤ.

Sampling
x[n]  x(t ) ST (t )

  x(t ) (t  nT )
n 

  x(nT ) (t  nT )
n 

• Sampling
− The spectrum of the discrete signal consists of

repetitions of the spectrum of the continuous
signal every 1/ΔΤ.
− The Nyquist criterion should be satisfied.
f (t )  F (  )

1  n 
f (nT )  f [n]  F (  ) 
T

n 
F  


T 

Nyquist theorem
1
 2  max
T

FT of a continuous signal
Oversampling
Critical sampling with

the Nyquist frequency
Undersampling
Aliasing appears
• Reconstruction (under correct sampling).

F ( )  F ( ) H ( )
 t 
f (t )  f (t ) * T sin c  
 T 
• Reconstruction
− Provided a correct sampling, the continuous

signal may be perfectly reconstructed by its
samples.

 (t  nT ) 
f (t )  
n 
f (nT ) sinc 
 n T 


• Under aliasing, the

reconstruction of
the continuous
signal not correct.

Aliased signal

21 The Discrete Fourier Transform
• The Fourier transform of a sampled (discrete)

signal is a continuous function of the frequency.


1  n 
F ( ) 
T

n 
F  


T 
• For a N-length discrete signal, taking N samples of
its Fourier transform at frequencies:
k
k  , k  0,1, .., N  1
N T
provides the discrete Fourier transform (DFT) of
the signal.
22
The Discrete Fourier Transform
(cont.)
• DFT pair of signal f [n] of length N.
N 1 2 nk
j
F [k ]   f [n]e N
, 0  k  N 1
n 0
N 1 2 nk
1
 F[k ]e
j
f [ n]  N
, 0  n  N 1
N n 0

23
(cont.)
• Property
– The DFT of a N-length f [n] signal is periodic

with period N.
F [k  N ]  F [k ]
– This is due to the periodicity of the complex
exponential:
2 n 2 n
 j  j
 wN   e  
n n

N N
wN e
 
24
(cont.)
• Property: sum of complex exponentials
1 N kn 1, k  rN , r 
 wN  
N n 0 0, otherwise
The proof is left as an exercise.
2
j
wN  e N

25
(cont.)
• DFT pair of signal f [n] of length N may be
expressed in matrix-vector form.

N 1
F [k ]   f [n]w , nk
N 0  k  N 1
n 0
1 N 1
f [n]   F [k ]wN ,
 nk
0  n  N 1
N n 0
2
j
wN  e N

26
(cont.)
F = Af
  w0  0 w  0 1
w  0 2
w 
0 N 1

 N N N N

 1 0 1 N 1 
 wN  w  1 1
w  w 
1 2
A  N N N 
 
 
 w N 1 0 w  w 
N 1 1 N 1 2
 wN  
N 1 N 1 
 N N N
f =  f [0], f [1],..., f [ N  1] , F =  F [0], F [1],..., F [ N  1]

T T

27
(cont.)
Example for N=4
1 1 1 1 
1  j 1 j 
A 
1 1 1 1
 
1 j 1  j 

28
(cont.)
The inverse DFT is then expressed by:
f = A -1F
* T
 0 
   wN  w  w   w  
0 0 1 0 2 0 N 1
N N N 
 1 0 1 N 1 

1  w  w 1 1
w   wN  
2

1
A 1 
1
N
 A 
* T
  N
N 
N N
 
  
  N 1 0 
  wN  w  w   wNN 1  
N 1 1 N 1 2 N 1

 N N

This is derived by the complex exponential sum

property.
29 Linear convolution
f [n]  {1, 2, 2}, h[n]  {1,  1}, N1  3, N 2  2

g[n]  f [n]* h[n]  
m 
f [m]h[n  m]
is of length N=N1+N2-1=4

30 Linear convolution (cont.)
f [n]  {1, 2, 2}, h[n]  {1,  1}, N1  3, N 2  2

g[n]  f [n]* h[n]  
m 
f [m]h[n  m]
f [ m] 1 2 2 g[ n]
n0 h[0  m] 1 1  1
n 1 h [1  m] 1 1  1
n2 h[2  m] 1 1  0
n3 h [3  m] 1 1  2
g[n]  {1, 1, 0,  2}
31 Circular shift
• Signal x[n] of length N.

• A circular shift ensures that the resulting
signal will keep its length N.
• It is a shift modulo N denoted by
x[(n  m) N ]  x[(n  m) mod N ]
• Example: x[n] is of length N=8.

x[(2) N ]  x[(2)8 ]  x[6]
x[(10) N ]  x[(10)8 ]  x[2]
32 Circular convolution
f [n]  {1, 2, 2}, h[n]  {1,  1}, N1  3, N2  2

g[n]=f [n]h[n]  
m 
f [m]h[(n  m) N ]
Circular shift modulo N
The result is of length N  max N1 , N2   3

33 Circular convolution (cont.)
f [n]  {1, 2, 2}, h[n]  {1,  1}, N1  3, N 2  2

g[n]=f [n]h[n]  
m 
f [m]h[(n  m) N ]
f [ m] 1 2 2 g[ n]
n  0 h [(0  m) N ] 1 1 1 1
n  1 h [(1  m) N ] 1 1 1
n  2 h [(2  m) N ] 1 1 0
g[n]  { 1, 1, 0,}
34 DFT and convolution
g[n]=f [n]h[n]  G[k ]  F [k ]H [k ]
• The property holds for the circular

convolution.
• In signal processing we are interested in
linear convolution.
• Is there a similar property for the linear
convolution?

35 DFT and convolution (cont.)
g[n]=f [n]h[n]  G[k ]  F [k ]H [k ]
• Let f [n] be of length N1 and h[n] be of length N2.
• Then g[n]=f [n]*h[n] is of length N1+N2-1.
• If the signals are zero-padded to length N=N1+N2-1
then their circular convolution will be the same as
their linear convolution:
g[n]  f [n]* h[n]  G[k ]  F [k ]H [k ]
Zero-padded signals
f [n]  {1, 2, 2}, h[n]  {1,  1}, N1  3, N 2  2
Zero-padding to length N=N1+N2-1 =4

f [n]  {1, 2, 2, 0}, h[n]  {1,  1, 0, 0}
f [ m] 1 2 2 0 g[ n]
h[(n  0)4 ] 0 0 1 1 0 0 1 1
h[(n  1) 4 ] 0 0 1 1 0 0 1
h[(n  2) 4 ] 0 0 1 1 0 0
h[(n  3)4 ] 0 0 1 1 2
The result is the same as the linear convolution.
Verification using DFT
1 1 1 1  1   5 
1  j 1 j   2  1  j 2 
F = Af       
1 1 1 1  2   1 
    
1 j 1  j  0   1  j 2
1 1 1 1   1  0 
1  j 1 j   1 1  j 
H = Ah       
1 1 1 1  0   2 
    
1 j 1  j   0  1  j 
G[k ]  F [k ] H [k ]
Element-wise multiplication
 5 0   0 
(1  j 2)  (1  j )  1  j 3
G = FH   
 1 2   2 
   
(1  j 2)  (1  j )  1  j3

Inverse DFT of the result
1 1 1 1   0   1
1 j 1  j  1  j 3  1
g = A G = A 
1 1 * T    
4 1 1 1 1  2   0
    
1  j 1 j  1  j 3  2
The same result as their linear convolution.

40 2D continuous signals
, x  x0 , y  y0
 ( x  x0 , y  y0 )  
 0 otherwise
 

 
f ( x, y ) ( x  x0 , y  y0 )dydx  f ( x0 , y0 )
Separable:  ( x  x0 , y  y0 )   ( x  x0 ) ( y  y0 )
The 2D impulse train is also separable:

 
S X Y ( x, y )  S X ( x) S Y ( y )     ( x  nX , y  nY )
n  m 

• The Fourier transform of a continuous 2D

signal f (x,y).
 
F (  , )  
 j 2 (  x  vy )
f ( x, y )e dydx
 
 
  F ( , )e d d 
j 2 (  x  vy )
f ( x, y ) 
 

• Example: FT of f (x,y)=δ(x)
y f (x,y)=δ(x)
 
F (  , )      j 2 (  x  vy )
( x ) e dydx x
 
 
  
 j 2 x  j 2 vy
 ( x ) e dx e dy
  ν F(μ,ν)=δ(ν)

e
 j 2 vy
 dy   ( ) μ


• Example: FT of f (x,y)=δ(x-y) f (x,y)=δ(x-y)
y
 
F (  , )      j 2 (  x  vy )
( x  y ) e dydx x
 

    j 2 vy
     ( x  y)e  j 2 x
dx e dy
    ν F(μ,ν)=δ(μ+ν)
 

 j 2 (   ) y
 e
 j 2 y  j 2 y
e dy  e dy μ
 
  (   )
sin(W ) sin( W )
f ( x, y)   PW / 2,W / 2 ( x, y)  F (  , )  W 2
(W ) ( W )
• 2D continuous convolution
 
f ( x, y )* h( x, y )  
 
f ( x   , y   )h( ,  )d d 
• We will examine the discrete convolution

in more detail.
• Convolution property
f ( x, y ) * h( x, y)  F (  , ) H (  , )
• 2D sampling is accomplished by
 
SX Y ( x, y )     ( x  nX , y  nY )
n  m 
• The FT of the sampled 2D signal consists of

repetitions of the spectrum of the 1D
continuous signal.

1 1  m n 
F (  , ) 
X Y

n 
F  
 X
,  
Y 

• The Nyquist theorem involves both the

horizontal and vertical frequencies.

1 1
 2max ,  2vmax
X Y
Over-sampled Under-sampled
49

Aliasing
50 Aliasing - Moiré Patterns
• Effect of sampling a scene with periodic or
nearly periodic components (e.g.
overlapping grids, TV raster lines and
stripped materials).
• In image processing the problem arises
when scanning media prints (e.g.
magazines, newspapers).
• The problem is more general than
sampling artifacts.
51 Aliasing - Moiré Patterns (cont.)
• Superimposed grid drawings (not digitized)
produce the effect of new frequencies not
existing in the original components.

• In printing industry the problem comes
when scanning photographs from the
superposition of:
• The sampling lattice (usually horizontal and
vertical).
• Dot patterns on the newspaper image.


• The printing
industry uses
halftoning to
cope with the
problem.
• The dot size is
inversely
proportional to
image intensity.

55 2D discrete convolution
m f [m,n] m h [m,n]
1 1
3 2 1 -1 n
n
 
g[m, n]  f [m, n]* h[m, n]    f [k , l ]h[m  k , n  l ]
k  l 
• Take the symmetric of one of the signals with

respect to the origin.
• Shift it and compute the sum at every position
[m,n].
56 2D discrete convolution (cont.)
n f [m,n] n h [m,n]
1 1
3 2 1 -1 m
m
l l
g [0,0]=0 g [1,1]=0
h [-k,-l] h [1-k,1-l]
3 2 3 2
k -1 1 k
-1 1
1 1
1 1
n f [m,n] n h [m,n]
1 1
3 2 1 -1 m
m
l l
g [2,2]=3 g [3,2]=-1
h [2-k,2-l] h [3-k,2-l]
-1 13 2 k
3 21
-1 k
1 1 1 1

n f [m,n] n h [m,n]
1 1
3 2 1 -1 m
m
n g[m,n]
3 5 2
3 -1 -3
m
M1+M2-1=3
N1+N2-1=2
59 The 2D DFT
• 2D DFT pair of image f [m,n] of size MxN.

M 1 N 1  km ln 
 j 2   
F [k , l ]    f [m, n]e M N
m0 n 0
M 1 N 1  km ln 
1 j 2   
f [m, n] 
MN
  F[k , l ]e
m0 n 0
M N
0  k  M  1 0  m  M  1
 , 
 0  l  N 1  0  n  N 1
60 The 2D DFT (cont.)
• All of the properties of 1D DFT hold.

• Particularly:
– Let f [m,n] be of size M1xN1 and h[m,n] of size
M2xN2 .
– If the signals are zero-padded to size (M1+M2-
1)x(N1+N2-1) then their circular convolution will
be the same as their linear convolution and:
g[m, n]  f [m, n]* h[m, n]  G[k , l ]  F [k , l ]H [k , l ]

Chapter 04a Frequency Filtering (Fundamentals)

Uploaded by

Document Information

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Chapter 04a Frequency Filtering (Fundamentals)

Uploaded by

Copyright:

Available Formats

Digital Image Processing

Filtering in the Frequency Domain

University of Ioannina - Department of Computer Science

Frequency: The number of times that a periodic

Webster’s New Collegiate Dictionary

C. Nikou – Digital Image Processing (E12)

Any function that periodically repeats itself can

C. Nikou – Digital Image Processing (E12)

Impulse train function

C. Nikou – Digital Image Processing (E12)

x[n]  x(t ) ST (t )

C. Nikou – Digital Image Processing (E12)

• The Fourier series expansion of a periodic

C. Nikou – Digital Image Processing (E12)

• The Fourier transform of a continuous

• Convolution property of the FT.

C. Nikou – Digital Image Processing (E12)

− The Fourier transform of the impulse train.

• It is also an impulse train in the frequency

C. Nikou – Digital Image Processing (E12)

x[n]  x(t ) ST (t )

C. Nikou – Digital Image Processing (E12)

− The spectrum of the discrete signal consists of

C. Nikou – Digital Image Processing (E12)

C. Nikou – Digital Image Processing (E12)

Critical sampling with

• Reconstruction (under correct sampling).

− Provided a correct sampling, the continuous

C. Nikou – Digital Image Processing (E12)

• Under aliasing, the

C. Nikou – Digital Image Processing (E12)

C. Nikou – Digital Image Processing (E12)

• The Fourier transform of a sampled (discrete)

signal is a continuous function of the frequency.

C. Nikou – Digital Image Processing (E12)

– The DFT of a N-length f [n] signal is periodic

The proof is left as an exercise.

C. Nikou – Digital Image Processing (E12)

expressed in matrix-vector form.

C. Nikou – Digital Image Processing (E12)

f =  f [0], f [1],..., f [ N  1] , F =  F [0], F [1],..., F [ N  1]

C. Nikou – Digital Image Processing (E12)

C. Nikou – Digital Image Processing (E12)

This is derived by the complex exponential sum

f [n]  {1, 2, 2}, h[n]  {1,  1}, N1  3, N 2  2

C. Nikou – Digital Image Processing (E12)

• Signal x[n] of length N.

• Example: x[n] is of length N=8.

f [n]  {1, 2, 2}, h[n]  {1,  1}, N1  3, N2  2

Circular shift modulo N

The result is of length N  max N1 , N2   3

C. Nikou – Digital Image Processing (E12)

g[n]=f [n]h[n]  G[k ]  F [k ]H [k ]

• The property holds for the circular

C. Nikou – Digital Image Processing (E12)

g[n]  f [n]* h[n]  G[k ]  F [k ]H [k ]

Zero-padding to length N=N1+N2-1 =4

C. Nikou – Digital Image Processing (E12)

Inverse DFT of the result

The same result as their linear convolution.

C. Nikou – Digital Image Processing (E12)

The 2D impulse train is also separable:

C. Nikou – Digital Image Processing (E12)

• The Fourier transform of a continuous 2D