ENG4BF3

Medical Image Processing

Image Enhancement in
Frequency Domain
 Image Enhancement

Original image Enhanced image

Enhancement: to process an image for more

suitable output for a specific application.

2
 Image Enhancement

• Image enhancement techniques:

 Spatial domain methods
 Frequency domain methods

• Spatial (time) domain techniques are techniques that

operate directly on pixels.

• Frequency domain techniques are based on modifying

the Fourier transform of an image.

3
 Fourier Transform: a review

• Basic ideas:
 A periodic function can be
represented by the sum of
sines/cosines functions of
different frequencies,
multiplied by a different
coefficient.
 Non-periodic functions can
also be represented as the
integral of sines/cosines
multiplied by weighing
function.

4
Joseph Fourier
(1768-1830)

Fourier was obsessed

with the physics of
heat and developed
the Fourier transform
theory to model heat-
flow problems.

5
 Fourier transform
basis functions

Approximating a
square wave as the
sum of sine waves.

6
 Any function can be written as the
sum of an even and an odd function

E ( x)  [ f ( x)  f ( x)] / 2

E(-x) = E(x)

O( x)  [ f ( x)  f ( x )] / 2

O(-x) = -O(x)


f ( x)  E ( x)  O( x)

7
 Fourier Cosine Series
Because cos(mt) is an even function, we can write an even
function, f(t), as:
  

F
1
  f(t)  m cos(mt)

m 0
where series Fm is computed as
  

Fm 


f (t ) cos(mt ) dt

Here we suppose f(t) is over the interval (–π,π).

8
 Fourier Sine Series

Because sin(mt) is an odd function, we can write any

odd function, f(t), as:
 


1
  f (t)  Fmsin(mt)

m 0

where the series F’m is computed as

  

  Fm 


f (t ) sin( mt ) dt

9
 Fourier Series

So if f(t) is a general function, neither even nor

odd, it can be written:

1  1 
f (t )  
 m 0
Fm cos(mt )  
 m 0
Fm sin(mt )

Even component Odd component

where the Fourier series is

Fm 
 f (t) cos(mt) dt Fm 
 f (t) sin(mt) dt

10
 The Fourier Transform
Let F(m) incorporates both cosine and sine series coefficients,
with the sine series distinguished by making it the imaginary
component:
F (m)  Fm  jFm'   f (t ) cos(mt )dt  j   f (t ) sin(mt )dt

Let’s now allow f(t) range from – to , we rewrite:


 f (t )  F (u )   f (t ) exp( j 2ut )dt


F(u) is called the Fourier Transform of f(t). We say that f(t)

lives in the “time domain,” and F(u) lives in the “frequency
domain.” u is called the frequency variable.
11
The Fourier Transform

u·x  1/M

12
d a

time domain frequency domain

d=a+b+ b time domain frequency domain

c
time domain frequency domain time domain frequency domain

13
 The Inverse Fourier Transform

We go from f(t) to F(u) by


Fourier
 f (t )  F (u )   f (t ) exp( j 2ut )dt

Transform

Given F(u), f(t) can be obtained by the inverse Fourier transform

 Inverse
1  F (u )  f (t )   F (u ) exp( j 2ut )du Fourier
 Transform

14
 Discrete Fourier Transform (DFT)
• A continuous function f(x) is discretized as:
{ f ( x0 ), f ( x0  x), f ( x0  2x),..., f ( x0  ( M  1)x)}

15
 Discrete Fourier Transform (DFT)

Let x denote the discrete values (x=0,1,2,…,M-1),

i.e.
f ( x)  f ( x0  xx)
then
{ f ( x0 ), f ( x0  x), f ( x0  2x),..., f ( x0  ( M  1)x)}

{ f (0), f (1), f (2),..., f ( M  1)}

16
 Discrete Fourier Transform (DFT)
• The discrete Fourier transform pair that applies to
sampled functions is given by:

M 1
1
F (u ) 
M
 f ( x) exp( j 2ux / M )
x 0
u=0,1,2,…,M-1

and
M 1
f ( x)   F (u ) exp( j 2ux / M ) x=0,1,2,…,M-1
u 0

17
 2-D Discrete Fourier Transform

• In 2-D case, the DFT pair is:

M 1 N 1
1
F (u, v) 
MN
 f ( x, y) exp( j 2 (ux / M  vy / N ))
x 0 y 0

u=0,1,2,…,M-1 and v=0,1,2,…,N-1

and:
M 1 N 1
f ( x, y )   F (u , v) exp( j 2 (ux / M  vy / N ))
u 0 v 0

x=0,1,2,…,M-1 and y=0,1,2,…,N-1

18
 Polar Coordinate Representation of FT

• The Fourier transform of a real function is generally

complex and we use polar coordinates:

F (u , v)  R (u , v)  j  I (u , v)
Polar coordinate
F (u , v)  F (u , v) exp( j (u , v))
Magnitude: F (u , v)  [ R 2 (u , v)  I 2 (u , v)]1/ 2
 I (u, v) 
Phase:  (u, v)  tan  1

 R(u, v) 
19
 Fourier Transform: shift
• It is common to multiply input image by (-1)x+y prior to
computing the FT. This shift the center of the FT to (M/2,N/2).

 f ( x, y )  F (u , v)
 f ( x, y )(1) x  y   F (u  M / 2, v  N / 2)

Shift

20
 Symmetry of FT

• For real image f(x,y), FT is conjugate symmetric:

F (u, v)  F (u, v) *

• The magnitude of FT is symmetric:

F (u , v)  F (u , v)

21
FT

IFT

22
IFT

IFT

23
The central part of FT, i.e.
the low frequency
components are responsible
for the general gray-level
appearance of an image.

The high frequency

components of FT are
responsible for the detail
information of an image.

24
Image Frequency Domain
(log magnitude)
v Detail

General
appearance

25
5% 10 % 20 % 50 %

26
Frequency Domain Filtering

27
 Frequency Domain Filtering

• Edges and sharp transitions (e.g., noise) in an

image contribute significantly to high-frequency
content of FT.
• Low frequency contents in the FT are
responsible to the general appearance of the
image over smooth areas.
• Blurring (smoothing) is achieved by attenuating
range of high frequency components of FT.

28
 Convolution in Time Domain
g(x,y)=h(x,y)f(x,y)
M 1 M 1
g( x , y )    h( x' , y' ) f ( x  x' , y  y' )
x'  0 y '  0

 f ( x , y )* h( x , y )
– f(x,y) is the input image
– g(x,y) is the filtered
– h(x,y): impulse response

29
 Convolution Theorem
Multiplication in
G(u,v)=F(u,v)H(u,v) Frequency Domain

g(x,y)=h(x,y)f(x,y) Convolution in
Time Domain

• Filtering in Frequency Domain with H(u,v) is

equivalent to filtering in Spatial Domain with f(x,y).

30
blue line = sum of 3 sinusoids (20,
50, and 80 Hz) + random noise
red line = sum of 3 sinusoids
without
withou noise

blue line = sum of 3 sinusoids after

filtering in time domain
1x average [1 1 1 1 1] / 5
blue line = sum of 3 sinusoids after
filtering in frequency domain
cut-off 90 Hz

31
 Algorithm Complexity

• We can compute the DFT directly using the formula

– An N point DFT would require N2 floating point multiplications per
output point
– Since there are N2 output points , the computational complexity of the
DFT is N4
– N4=4x109 for N=256
– Bad news! Many hours on a workstation

32
 Algorithm Complexity

• The FFT algorithm was developed in the 60’s for seismic

exploration
• Reduced the DFT complexity to 2N2log2N
– 2N2log2N~106 for N=256
– A few seconds on a workstation

33
 Examples of Filters

Frequency
domain

Gaussian lowpass filter Gaussian highpass filter

Spatial
domain

34
 Ideal low-pass filter (ILPF)

1 D(u, v)  D0
H (u, v)  
0 D(u, v)  D0
D(u , v)  [(u  M / 2)  (v  N / 2) ]
2 2 1/ 2

(M/2,N/2): center in frequency domain

D0 is called the cutoff frequency.

35
 Shape of ILPF

Frequency domain

h(x,y)

Spatial domain

36
 FT

ringing
and
Ideal in frequency blurring
domain means
non-ideal in
spatial domain,
vice versa.

37
 Butterworth Lowpass Filters (BLPF)
• Smooth transfer function, 1
no sharp discontinuity, H (u, v)  2n
 D(u, v) 
no clear cutoff 1  
 D0 
frequency.

1
2

38
 Butterworth Lowpass Filters (BLPF)

n=1 n=2 n=5 n=20

h(x)

1
H (u, v)  2n
 D(u , v) 
1  
 D 0 
39
No serious
ringing
artifacts

40
 Gaussian Lowpass Filters (GLPF)
• Smooth transfer function,
2
smooth impulse 
D ( u ,v )

response, no ringing H (u , v)  e 2 D 20

41
 GLPF

Frequency
domain

Gaussian lowpass filter

Spatial
domain

42
No ringing
artifacts

43
Examples of Lowpass Filtering

44
 Examples of Lowpass Filtering

Low-pass filter H(u,v)

Original image and its FT Filtered image and its FT

45
 High-pass Filters
• Hhp(u,v)=1-Hlp(u,v)

1 D(u, v)  D0
• Ideal: H (u, v)  
0 D(u, v)  D0

1
• Butterworth: | H (u, v) | 
2
2n
 D0 
1  
 D (u , v ) 

• Gaussian:  D 2 ( u ,v ) / 2 D02
H (u , v)  1  e

46
47
 High-pass Filters

h(x,y)

48
 Ideal High-pass Filtering

ringing artifacts

49
Butterworth High-pass Filtering

50
Gaussian High-pass Filtering

51
 Gaussian High-pass Filtering

Gaussian filter H(u,v)

Original image

Filtered image and its FT

52
 Laplacian in Frequency Domain

 2 f ( x , y )  2 f ( x, y )
[  ]   2
 2
(u v ) F (u, v)
x 2
y 2

H1 (u , v)   2

(u v )2

Frequency
Spatial domain
 f 2
 f domain
2
 f  2  2
2
Laplacian operator
x y

53
54
 Subtract Laplacian from the Original
Image to Enhance It

enhanced Original Laplacian

image image output

Spatial 
g ( x , y ) f ( x, y )   2
f ( x, y )
domain

Frequency
G (u, v)  F (u, v)  (u 2  v 2 ) F (u , v)
domain
new operator H 2 (u , v)  1  (u 2  v 2 )  1  H 1 (u , v)

Laplacian

55
f 2 f

f  f
2

56
 Unsharp Masking, High-boost Filtering

• Unsharp masking: fhp(x,y)=f(x,y)-flp(x,y)

• Hhp(u,v)=1-Hlp(u,v)
One more
parameter to
• High-boost filtering: adjust the
enhancement
fhb(x,y)=Af(x,y)-flp(x,y)
• fhb(x,y)=(A-1)f(x,y)+fhp(x,y)
• Hhb(u,v)=(A-1)+Hhp(u,v)
57
58
End of Lecture

59