Image Enhancement in the

Frequency Domain

Spring 2006, Jen-Chang Liu

 Outline
 Introduction to the Fourier Transform and
Frequency Domain
 Magnitude of frequencies
 Phase of frequencies
 Fourier transform and DFT
 Filtering in the frequency domain
 Smoothing Frequency Domain Filters
 Sharpening Frequency Domain Filters
 Homomorphic Filtering
 Implementation of Fourier transform
 Background
 1807, French math. Fourier
 Any function that periodically repeats itself can
be expressed as the sum of of sines and/or
cosines of different frequencies, each multiplied
by a different coefficient (Fourier series)
Periodic function

f(t) = f(t+T), T: period (sec.)

1/T: frequency (cycles/sec.)
 Frequency Weight
f1 w1
f2 w2

f3 w3

f4 w4

Periodic function f
 How to measure weights?
 Assume f1 , f2 , f3 , f4 are known

f (t )  w1 f1 (t )  w2 f 2 (t )  w3 f 3 (t )  w4 f 4 (t )
 How to measure w1 , w2 , w3 , w4 ?

min      2
[ f (t ) ( w f
1 1 (t ) w f
2 2 (t ) w f
3 3 (t ) w f
4 4 (t ))] dt

Minimize squared error

 Minimize MSE calculation
min F ( w1 , w2 , w3 , w4 )
  [ f (t )  ( w1 f1 (t ) w2 f 2 (t )  w3 f 3 (t )  w4 f 4 (t ))] dt
2

  ( f  w f w f  w f  w f
2 2
1 1
2 2
2 2
2 2
3 3
2 2
4 4
2

 2 w1 ff1  2 w1w2 f 2 f1  2 w1w3 f 3 f1  2 w1w4 f 4 f1

 2 w2 ff 2  2 w2 w3 f 3 f 2  2 w2 w4 f 4 f 2
 2 w3 ff3  2 w4 w3 f 4 f 3
 2 w4 ff 4 ) dt
 Orthogonal condition 正交

 f1 and f2 are orthogonal if

f1 , f 2   f1 (t ) f 2 (t )dt  0

 f1 , f2 , f3 , f4 are orthogonal to each other

F ( w1 , w2 , w3 , w4 )
  ( f 2  w12 f12 w22 f 22  w32 f 32  w42 f 42
 2w1 ff1 - 2w2 ff 2 - 2w3 ff3 - 2w4 ff 4 ) dt
 Minimization calculation
 To satisfy min F ( w1 , w2 , w3 , w4 )
  ( f 2  w12 f12 w22 f 22  w32 f 32  w42 f 42
 2w1 ff1 - 2w2 ff 2 - 2w3 ff3 - 2w4 ff 4 ) dt
F
We have  2w1  f1  2 ff1  0
2

w1

=> w1 
 ff 1

f , f1 Recall in linear
algebra: projection
f
2 2
1 f1
 Weight = Projection magnitude
 Represent input f(x) with another basis
functions
Functional space
Vector space

v f
projection

(1,0) f1
 Summary 1
 A function f can be written as sum of f1 , f2 ,
f3 , …
f (t )   wi f i (t )
i

If f1 , f2 , f3 , … are orthogonal to each other

f , fi
Weight (magnitude) wi  2
fi
 Summary 1: sine, cosine bases
 Let f1 , f2 , f3 , … carry frequency information
 Let them be sines and cosines
  if n  k  1
  n, k:integers
 cos( nt ) cos( kt) dt  2 if n  k  0
 0 otherwise

  if n  k  1
 sin( nt ) sin( kt)dt   0 otherwise

  sin( nt ) cos(kt)dt  0

for all integers n, k

=> They all satisfy orthogonal conditions

Summary 1: orthogonal
 Fourier series
 For f (t ), 0  t  2 (Assume periodic outside)


f (t )  a0   ak cos( kt)  bk sin( kt)
k 1

a0 , a1, b1, a2 , b2 , a3 , b3 ,...

DC
頻率=1 頻率=2 頻率=3
 Outline
 Introduction to the Fourier Transform and
Frequency Domain
 Magnitude of frequencies
 Phase of frequencies
 Fourier transform and DFT
 Filtering in the frequency domain
 Smoothing Frequency Domain Filters
 Sharpening Frequency Domain Filters
 Homomorphic Filtering
 Implementation of Fourier transform
 相位
Correlation with different phase
Weight calculation f , f1

w1  2
f1

f 相關係數

f1
 Correlation with different phase
(cont.)
Weight calculation f , f1

w1  2
 f1

f 相關係數?

f1
 Deal with phase: method 1
 For example, expand f(t) over the cos(wt)
basis function
 Consider different phases cos( wt  )
Corr ( )   f (t ) cos( wt   )dt

Corr()


0 0 2

Problem: weight(w, )
 Deal with phase: method 2
 Complex exponential as basis j  1
j
e  cos(t )  j sin( t )
jt

sin  1
 real
With frequency w: cos
e jwt
 cos( wt )  j sin( wt )
Advantage:
Derive magnitude and phase  simultaneously
 Deal with phase 2: example
 Input f (t )  cos(t   )
w  cos(t   ), e jt
2
 cos(t   )e  jt
dt
0
2 2
  cos(t ) cos( )e dt   sin( t ) sin(  )e jt dt
 jt
0 0
2 2
 cos( ) cos(t ) cos(t )dt  sin(  ) j  sin( t ) sin( t )dt
0 0
j
  cos( )  j sin(  )  e phase

magnitude
 Fourier series with phase
 For f (t ), 0  t  2 (Assume periodic outside)

 
f (t )   wk e jkt
  wk (cos kt  j sin kt)
k 0 k 0

w0 , w1 , w2 , w3 ,... Complex weight

DC
頻率k=1 k=2 k=3
 Outline
 Introduction to the Fourier Transform and
Frequency Domain
 Magnitude of frequencies
 Phase of frequencies
 Fourier transform and DFT
 Filtering in the frequency domain
 Smoothing Frequency Domain Filters
 Sharpening Frequency Domain Filters
 Homomorphic Filtering
 Implementation of Fourier transform
 Fourier transform
 Functions that are not periodic can be
expressed as the integral of sines and/or
cosines multiplied by a weighting functions
 Frequency up to infinity
 Perfect reconstruction
Functions -- Fourier transform

Operation in frequency domain

without loss of information
 1-D Fourier Transform
 Fourier transform F(u) of a continuous
function f(x) is:
Inverse transform:


j 2ux j  1
f ( x)  F (u )e du


Forward Fourier transform:


F (u )  
f ( x)e  j 2ux dx
 2-D Fourier Transform
 Fourier transform F(u,v) of a continuous
function f(x,y) is:
Inverse transform:
 
f ( x, y)    F (u, v)e j 2 (uxvy ) dudv
  

x u

y F
v
 Future development
 1950, fast Fourier transform (FFT)
 Revolution in the signal processing
 Discrete Fourier transform (DFT)
 For digital computation
 1-D Discrete Fourier Transform
 f(x), x=0,1,…,M-1 . discrete function
 F(u), u=0,1,…,M-1. DFT of f(x)
Inverse transform:
M 1 j 2
u

 F (u)e
x
f ( x)  M

u 0

Forward discrete Fourier transform:

M 1  j 2
u
1
 f ( x)e
x
F (u )  M
M x 0
 Frequency Domain 頻率域
 Where is the frequency domain?
j
Euler’s formula:
sin  1
e j  cos   j sin  
cos
M 1
1 u u
F (u ) 
M

x 0
f ( x)[cos 2
M
x  j sin 2
M
x]

F(u) frequency

u
Fourier
transform
 Physical analogy
 Mathematical frequency splitting
 Fourier transform
 Physical device
 Galss prism 三稜鏡
 Split light into frequency components
 F(u) Complex quantity?
imaginary
 Polar coordinate
F (u)  R(u)  jI (u) m
 real
 j ( u )
 F (u) e

F (u)  [ R (u)  I (u)]

2 2 1/ 2
magnitude

I (u )
 (u )  tan [ 1
] phase
R (u )

P(u )  F (u )  R (u )  I (u )
2 2 2
Power spectrum
 Some notes about sampling in
time and frequency axis
 Time index

f ( x)  f ( x0  xx) x  0,1,..., M  1

[ x0 , x0  x , ...., x0  (M  1)x]

 Frequency index
F (u )  F (uu ) u  0,1,..., M  1
 Also follow reciprocal property
u  1
x
 Extend to 2-D DFT from 1-D
 2-D: x-axis then y-axis

M 1 N 1 j 2 (
u
x
v

 F (u, v)e
y)
f ( x, y )  M M

u 0 v 0

N 1 M 1  j 2 (
u
x
v
1
  f ( x, y ) e
y)
F (u , v)  M M
MN y 0 x 0
 Complex Quantities to Real
Quantities
 Useful representation
F (u, v)  [ R (u, v)  I (u, v)]
2 2 1/ 2
magnitude

I (u, v)
 (u, v)  tan [ 1
] phase
R(u , v)
Power spectrum
P(u, v)  F (u, v)  R (u, v)  I (u, v)
2 2 2
DFT: example

log(F)
 Properties in the frequency
domain
 Fourier transform works globally
 No direct relationship between a specific
components in an image and frequencies
 Intuition about frequency
 Frequency content

 Rate of change of gray levels in an image

+45,-45 degree

artifacts