CS589-04 Digital Image Processing

Lecture 3. Filtering in the

Frequency Domain

Spring 2008
New Mexico Tech
 Outline

► Fourier Transform

► Filtering in Fourier Transform Domain

04-Dec-11 2
 Fourier Series and Fourier Transform: History
► Jean Baptiste Joseph Fourier, French mathematician and physicist
(03/21/1768-05/16/1830) http://en.wikipedia.org/wiki/Joseph_Fourier

Permanent
Orphaned: at nine
Secretary of the
French Academy of
Egyptian expedition Sciences: 1822
with Napoleon I:
1798 Théorie analytique
Governor of Lower de la chaleur :
Egypt 1822

(The Analytic
Theory of Heat)

04-Dec-11 3
 Fourier Series and Fourier Transform: History

► Fourier Series
Any periodic function can be expressed as the sum of sines
and /or cosines of different frequencies, each multiplied by
a different coefficients

► Fourier Transform
Any function that is not periodic can be expressed as the
integral of sines and /or cosines multiplied by a weighing
function

04-Dec-11 4
 Fourier Series: Example

04-Dec-11 5
 Preliminary Concepts

j = −1, a complex number

C = R + jI
the conjugate
C* = R - jI

| C |= R 2 + I 2 and θ = arctan( I / R )
C =| C | (cos θ + j sin θ )
Using Euler's formula,
C =| C | e jθ

04-Dec-11 6
 Fourier Series

A function f (t ) of a continuous variable t that is periodic

with period, T , can be expressed as the sum of sines and
cosines multiplied by appropriate coefficients
∞ 2π n

∑ce
j t
f (t ) = n
T

n =−∞

where
2π n
1 T /2 −j t
cn = ∫ f (t )e T
dt for n = 0, ±1, ±2,...
T −T /2

04-Dec-11 7
 Impulses and the Sifting Property (1)
A unit impulse of a continuous variable t located
at t =0, denoted δ (t ), defined as
∞ if t = 0
δ (t ) = 
0 if t ≠ 0
and is constrained also to satisfy the identity
∞
∫ −∞
δ (t )dt = 1
∞

The sifting property ∫ −∞

f (t )δ (t − t0 )dt = f (t0 )
∞
∫−∞
f (t )δ (t )dt = f (0)
04-Dec-11 8
 Impulses and the Sifting Property (2)
A unit impulse of a discrete variable x located
at x =0, denoted δ ( x), defined as
1 if x = 0
δ ( x) = 
0 if x ≠ 0
and is constrained also to satisfy the identity
∞

∑ δ ( x) = 1
x =−∞
∞
The sifting property ∑
x =−∞
f ( x)δ ( x − x0 ) = f ( x0 )
∞

∑
x =−∞
f ( x)δ ( x) = f (0)
04-Dec-11 9
 Impulses and the Sifting Property (3)

impulse train s∆T (t ),

∞
s∆T (t ) = ∑ δ (t − n∆T )
n =−∞

04-Dec-11 10
 Fourier Transform: One Continuous Variable

The Fourier Transform of a continous function f (t )

∞
F ( µ ) = ℑ{ f (t )} = ∫ f (t )e − j 2πµt dt
−∞

The Inverse Fourier Transform of F ( µ )

∞
f (t ) = ℑ {F ( µ )} = ∫ F ( µ )e j 2πµt d µ
−1
−∞

04-Dec-11 11
 Fourier Transform: One Continuous Variable

∞ W /2
F (µ ) = ∫ f (t )e − j 2πµ t
dt = ∫ Ae − j 2πµt dt
−∞ −W /2

−A − j 2πµ t W /2 A
= e  = e jπµW − e − jπµW 
j 2πµ −W /2 j 2π W
sin(πµW )
= AW
(πµW )

04-Dec-11 12
 Fourier Transform: Impulses

The Fourier transform of a unit impulse located at the origin:

∞
F ( µ ) = ∫ δ (t )e − j 2πµt dt
−∞

= e − j 2πµ 0
=1
The Fourier transform of a unit impulse located at t = t0 :
∞
F ( µ ) = ∫ δ (t − t0 )e − j 2πµt dt
−∞

= e − j 2πµt0
=cos(2πµ t0 ) − j sin (2πµ t0 )

04-Dec-11 13
 Fourier Transform: Impulse Trains
∞
Impulse train s∆T (t ), s∆T (t ) = ∑ δ (t − n∆T )
n =−∞

The Fourier series:

∞ 2π n

∑c
j t
s∆T (t ) = n e ∆T

n =−∞

where
2π n
1 ∆T /2 −j t
cn =
∆T ∫−∆T /2
s∆T (t )e ∆T
dt

04-Dec-11 14
 Fourier Transform: Impulse Trains

 j 2∆π1Tn t ∆T /2 ∞ j 2∆−πTjn2∆tπTn t− j 2πµt 1 ∆T /2∞ j 2∆πT−nj t2∆πTn−t j 2πµt

ℑcn e= ∫−∆=

∆T

T /2 ∫−∞ ∆T −∆∫T −∞
s∆T (et )e e dt = dt ∫= δe(t )e e dt dt
/2

1 0 1
= e = n )t
∞ − j 2π ( µ −
∆T ∆∆TT n
=∫ e dt = δ ( µ − )
−∞ ∆T
 2π n n  ∞
2π n n
ℑ δ ( µj −∆T t ) 1= ∫ δ ( µj − t )e j 2πµt du
−∞
1 ∞
s∆T (t ) = ∑c n e ∆T=  ∑ −∞ e ∆T ∆T
n =−∞2π nt ∆T n =−∞
j
= e ∆T

04-Dec-11 15
 Fourier Transform: Impulse Trains
Let S ( µ ) denote the Fourier transform of the
periodic impulse train S∆T (t )
2π n
 1 ∞

∑e
j t
S ( µ ) = ℑ{S ∆T (t )} = ℑ  ∆T

 ∆T n =−∞ 
1  ∞ j 2∆πTn t 
= ℑ ∑ e 
∆T  n =−∞ 
1 ∞ n
= ∑
∆T n =−∞
δ (µ −
∆T
)

04-Dec-11 16
 Fourier Transform and Convolution
The convolution of two functions is denoted
by the operator
∞
f (t ) h(t ) = ∫ f (τ )h(t − τ )dτ
−∞
∞ ∞
ℑ{ f (t ) h(t )} = ∫ ∫ f (τ )h(t − τ )dτ e− j 2πµt dt

−∞ 
 −∞ 
∞ ∞
= ∫ f (τ ) ∫ h(t − τ )e − j 2πµt dt dτ

−∞  −∞ 
∞
=∫ f (τ )  H ( µ )e− j 2πµτ dτ
−∞
∞
=H ( µ ) ∫ f (τ )e− j 2πµτ dτ
−∞

=H ( µ ) F ( µ )
04-Dec-11 17
 Fourier Transform and Convolution

Fourier Transform Pairs

f (t ) h(t ) ⇔ H ( µ ) F ( µ )

f (t )h(t ) ⇔ H ( µ ) F ( µ )

04-Dec-11 18
 Fourier Transform of Sampled Functions

fɶ (t ) = f (t ) s∆T (t )
∞
= ∑
n =−∞
f (t )δ (t − n∆T )

04-Dec-11 19
 Fourier Transform of Sampled Functions

{ }
Fɶ ( µ ) = ℑ fɶ (t ) = ℑ{ f (t ) s∆T (t )} = F ( µ ) ? S ( µ )
∞
Fɶ ( µ ) = F1( µ )∞ S ( µ ) = ∫n F (τ ) S ( µ − τ )dτ
S (µ ) = ∑
∆1T n =−∞
∞
δ ( µ −
∞ ∆T
−∞
)
n
= ∫ F (τ ) ∑ δ ( µ − τ − )dτ
∆T −∞
n =−∞ ∆T
1 ∞ ∞ n
= ∑
∆T n =−∞ ∫ −∞
F (τ )δ ( µ − τ −
∆T
)dτ

1 ∞ n
= ∑
∆T n =−∞
F (µ −
∆T
)

04-Dec-11 20
 Question

The Fourier transform of the

sampled function (shown in the
following figure) is

1. Continuous

2. Discrete

04-Dec-11 21
 Fourier Transform of Sampled Functions
► A bandlimited signal is a signal whose Fourier transform
is zero above a certain finite frequency. In other words, if
the Fourier transform has finite support then the signal is
said to be bandlimited.

An example of a simple bandlimited signal is a sinusoid of

the form,

x(t ) = sin(2π ft + θ )

04-Dec-11 22
 Fourier Transform of Sampled Functions

− µ max µmax
Over-sampling
1
> 2 µmax
Fɶ ( µ ) = ∆T
∞
1 n
∆T
∑ F (µ −
∆T
) Critically-sampling
n =−∞ 1
= 2 µmax
∆T
under-sampling
1
< 2µmax
04-Dec-11 ∆T 23
 Nyquist–Shannon sampling theorem

► We can recover f (t ) from its sampled version if we can

isolate a copy of F ( µ ) from the periodic sequence of copies
of this function contained in Fɶ ( µ ) , the transform of the
sampled function fɶ (t )

1
► Sufficient separation is guaranteed if > 2 µmax
∆T
Sampling theorem: A continuous, band-limited function
can be recovered completely from a set of its samples if
the samples are acquired at a rate exceeding twice the
highest frequency content of the function
04-Dec-11 24
 Nyquist–Shannon sampling theorem

?
∞
f (t ) = ∫ F ( µ )e j 2πµt d µ
−∞

04-Dec-11 25
 Aliasing

If a band-limited function is sampled at a rate that is less

than twice its highest frequency?

The inverse transform will yield a corrupted function. This

effect is known as frequency aliasing or simply as
aliasing.

04-Dec-11 26
 Aliasing

04-Dec-11 27
 Aliasing

04-Dec-11 28
 Function Reconstruction from Sampled Data

F ( µ ) = H ( µ ) Fɶ ( µ )

f (t ) = ℑ −1
{F ( µ )}
=ℑ −1
{H (µ ) Fɶ (µ )}
= h(t ) fɶ (t )
∞
f (t ) = ∑ f (n∆T )sinc [ (t − n∆T ) / n∆T ]
n =−∞
04-Dec-11 29
 The Discrete Fourier Transform (DFT) of One
Variable

M −1
F ( µ ) = ∑ f ( x)e − j 2πµ x / M , µ =0,1,..., M − 1
x =0

M −1
1
f ( x) =
M
∑ F
µ =0
( µ ) e j 2πµ x / M
, x = 0,1, 2,..., M − 1

04-Dec-11 30
 2-D Impulse and Sifting Property: Continuous
∞ if t = z = 0
The impulse δ (t , z ), δ (t , z ) = 
0 otherwise
∞ ∞
and ∫ ∫ −∞ −∞
δ (t , z )dtdz = 1

The sifting property

∞ ∞
∫ ∫
−∞ −∞
f (t , z )δ (t , z )dtdz = f (0, 0)
and
∞ ∞
∫ ∫
−∞ −∞
f (t , z )δ (t − t0 , z − z0 )dtdz = f (t0 , z0 )

04-Dec-11 31
 2-D Impulse and Sifting Property: Discrete

1 if x = y = 0
The impulse δ ( x, y ), δ ( x, y ) = 
0 otherwise

The sifting property

∞ ∞

∑∑
x =−∞ y =−∞
f ( x, y )δ ( x, y ) = f (0, 0)

and
∞ ∞

∑∑
x =−∞ y =−∞
f ( x, y )δ ( x − x0 , y − y0 ) = f ( x0 , y0 )

04-Dec-11 32
 2-D Fourier Transform: Continuous

∞ ∞
F ( µ ,ν ) = ∫ ∫ f (t , z )e − j 2π ( µ t +ν z )
dtdz
−∞ −∞

and
∞ ∞
f (t , z ) = ∫ ∫ f ( µ ,ν )e j 2π ( µ t +ν z )
d µ dν
−∞ −∞

04-Dec-11 33
 2-D Fourier Transform: Continuous
∞ ∞
F ( µ ,ν ) = ∫ ∫ f (t , z )e − j 2π ( µt +ν z ) dtdz
−∞ −∞
T /2 Z /2
=∫ ∫ Ae − j 2π ( µt +ν z ) dtdz
−T /2 − Z /2

 sin(πµT )   sin(πν T ) 
= ATZ    πν T 
 πµ T  

04-Dec-11 34
 2-D Sampling and 2-D Sampling Theorem

2 − D impulse train:
∞ ∞
s∆T ∆Z (t , z ) = ∑ ∑ δ (t − m∆T , z − n∆Z )
m =−∞ n =−∞

04-Dec-11 35
 2-D Sampling and 2-D Sampling Theorem

Function f (t , z ) is said to be band-limited if its Fourier transform

is 0 outside a rectangle established by the intervals [-µmax ,µmax ]
and [-ν max ,ν max ], that is
F ( µ ,ν ) = 0 for | µ |≥ µmax and | ν |≥ ν max

Two-dimensional sampling theorem:

A continuous, band-limited function f (t , z ) can be recovered with
no error from a set of its samples if the sampling intervals are
1 1
∆T< and ∆Z<
2 µmax 2ν max
04-Dec-11 36
 2-D Sampling and 2-D Sampling Theorem

04-Dec-11 37
 Aliasing in Images: Example
In an image system, the
number of samples is fixed at
96x96 pixels. If we use this
system to digitize checkerboard
patterns …

Under-sampling

04-Dec-11 38
 Aliasing in Images: Example

Re-sampling

04-Dec-11 39
 Aliasing in Images: Example

Re-sampling

04-Dec-11 40
 Moiré patterns

► Moiré patterns are often an undesired artifact of images

produced by various digital imaging and computer graphics
techniques
e. g., when scanning a halftone picture or ray tracing a
checkered plane. This cause of moiré is a special case of
aliasing, due to under-sampling a fine regular pattern

http://en.wikipedia.org/wiki/Moiré_pattern

04-Dec-11 41
 Moiré patterns

04-Dec-11 42
 Moiré patterns

04-Dec-11 43
 Moiré patterns
A moiré pattern
formed by
incorrectly down-
sampling the
former image

04-Dec-11 44
 2-D Discrete Fourier Transform and Its
Inverse
DFT:
M −1 N −1 − j 2π ( µ x / M +ν y / N )

F ( µ ,ν ) = ∑ ∑ f ( x, y )e
x =0 y =0

µ = 0,1, 2,..., M − 1;ν = 0,1, 2,..., N − 1;

f ( x, y ) is a digital image of size M × N.

IDFT:
M −1 N −1 j 2π ( µ x / M +ν y / N )
1
f ( x, y ) =
MN
∑ ∑ F (µ ,ν )e
x =0 y =0

04-Dec-11 45
 Properties of the 2-D DFT
relationships between spatial and frequency intervals

Let ∆T and ∆Z denote the separations between samples,

then the seperations between the corresponding discrete,
frequency domain variables are given by
1
∆µ =
M ∆T
1
and ∆ν =
N ∆Z

04-Dec-11 46
 Properties of the 2-D DFT
translation and rotation

f ( x, y )e j 2π ( µ0 x / M +ν 0 y / N ) ⇔ F ( µ − µ0 ,ν −ν 0 )
and
− j 2π ( µ x0 / M +ν y0 / N )
f ( x - x0 , y - y0 ) ⇔ F ( µ ,ν )e

Using the polar coordinates

x = r cos θ y=rsinθ µ =ω cosϕ ν =ω sin ϕ
results in the following transform pair:
f (r ,θ + θ 0 ) ⇔ F (ω , ϕ + θ 0 )
04-Dec-11 47
 Properties of the 2-D DFT
periodicity

2 − D Fourier transform and its inverse are infinitely periodic

F ( µ ,ν ) = F ( µ + k1M ,ν ) = F ( µ ,ν + k2 N ) = F ( µ + k1M ,ν + k2 N )

f ( x, y ) = f ( x + k1M , y ) = f ( x, y + k2 N ) = f ( x + k1M , y + k2 N )

f ( x ) e j 2π ( µ0 x / M ) ⇔ F ( µ − µ 0 )
µ0 = M / 2, f ( x)(−1) x ⇔ F ( µ − M / 2)

f ( x, y )(−1) x + y ⇔ F ( µ − M / 2,ν − N / 2)

04-Dec-11 48
 Properties of the 2-D DFT
periodicity

04-Dec-11 49
 Properties of the 2-D DFT
Symmetry

04-Dec-11 50
 Properties of the 2-D DFT
Fourier Spectrum and Phase Angle

2-D DFT in polar form

F (u , v) =| F (u , v) | e jφ (u ,v )
Fourier spectrum
1/ 2
| F (u , v) |=  R (u , v) + I (u , v) 
2 2

Power spectrum
P (u, v) =| F (u , v) |2 = R 2 (u , v) + I 2 (u , v)
Phase angle
 I (u , v) 
φ (u,v)=arctan  
 R (u , v ) 

04-Dec-11 51
04-Dec-11 52
04-Dec-11 53
 Example: Phase Angles

04-Dec-11 54
Example: Phase Angles and The Reconstructed

04-Dec-11 55
 2-D Convolution Theorem
1-D convolution
M −1
f ( x ) h( x ) = ∑ f ( m)h( x − m)
m =0

2-D convolution
M −1 N −1
f ( x, y ) h( x, y ) = ∑ ∑ f ( m, n ) h ( x − m, y − n )
m =0 n =0

x = 0,1, 2,..., M − 1; y = 0,1, 2,..., N − 1.

f ( x , y ) h ( x , y ) ⇔ F ( u, v ) H ( u, v )
f ( x , y )h ( x , y ) ⇔ F (u, v ) H ( u, v )
04-Dec-11 56
 An Example of Convolution

Mirroring h
about the
origin

Translating
the mirrored
function by x

Computing the
sum for each
x
04-Dec-11 57
 An Example of Convolution

It causes the
wraparoun
d error

It can be
solved by
appending
zeros

04-Dec-11 58
 Zero Padding

► Consider two functions f(x) and h(x) composed of A and B

samples, respectively

► Append zeros to both functions so that they have the same

length, denoted by P, then wraparound is avoided by
choosing

P ≥A+B-1

04-Dec-11 59
 Zero Padding

► Let f(x,y) and h(x,y) be two image arrays of sizes A×B and
C×D pixels, respectively. Wraparound error in their
convolution can be avoided by padding these functions
with zeros

 f ( x, y ) 0 ≤ x ≤ A -1 and 0 ≤ y ≤ B -1
f p ( x, y ) = 
 0 A ≤ x ≤ P or B ≤ y ≤ Q

 h ( x, y ) 0 ≤ x ≤ C -1 and 0 ≤ y ≤ D -1
h p ( x, y ) = 
 0 C ≤ x ≤ P or D ≤ y ≤ Q

Here P ≥ A + C − 1; Q ≥ B + D − 1
04-Dec-11 60
 Summary

04-Dec-11 61
 Summary

04-Dec-11 62
 Summary

04-Dec-11 63
 Summary

04-Dec-11 64
 The Basic Filtering in the Frequency Domain

Why is the spectrum at

almost ±45 degree stronger
than the spectrum at other
directions?

04-Dec-11 65
 The Basic Filtering in the Frequency Domain

► Modifying the Fourier transform of an image

► Computing the inverse transform to obtain the processed

result

g ( x, y ) = ℑ−1{H (u , v) F (u , v)}

F (u , v) is the DFT of the input image

H (u , v) is a filter function.

04-Dec-11 66
 The Basic Filtering in the Frequency Domain

► In a filter H(u,v) that is 0 at the center of the transform

and 1 elsewhere, what’s the output image?

04-Dec-11 67
 The Basic Filtering in the Frequency Domain

04-Dec-11 68
 The Basic Filtering in the Frequency Domain

04-Dec-11 69
 Zero-Phase-Shift Filters
−1
g ( x, y ) = ℑ {H (u, v) F (u, v)}
F (u , v) = R(u , v) + jI (u , v)
−1
g ( x, y ) = ℑ [ H (u, v) R(u, v) + jH (u, v) I (u, v)]
Filters affect the real and imaginary parts equally,
and thus no effect on the phase.
These filters are called zero-phase-shift filters

04-Dec-11 70
 Examples: Nonzero-Phase-Shift Filters

Even small
Phasechanges
angle is in the phase angle can ishave
Phase angle
dramaticmultiplied
(usually by undesirable) effects on the
multiplied by filtered
output 0.5 0.5

04-Dec-11 71
 Summary:
Steps for Filtering in the Frequency Domain
1. Given an input image f(x,y) of size MxN, obtain the
padding parameters P and Q. Typically, P = 2M and Q = 2N.

2. Form a padded image, fp(x,y) of size PxQ by

appending the necessary number of zeros to f(x,y)

3. Multiply fp(x,y) by (-1)x+y to center its transform

4. Compute the DFT, F(u,v) of the image from step 3

5. Generate a real, symmetric filter function*, H(u,v), of

size PxQ with center at coordinates (P/2, Q/2)
*generate from a given spatial filter, we pad the spatial filter, multiply the expadded
array by (-1)x+y, and compute the DFT of the result to obtain a centered H(u,v).
04-Dec-11 72
 Summary:
Steps for Filtering in the Frequency Domain
6. Form the product G(u,v) = H(u,v)F(u,v) using array
multiplication

7. Obtain the processed image

{ }
g p ( x, y ) = real  ℑ−1 [G (u, v) ] (−1) x + y

8. Obtain the final processed result, g(x,y), by extracting

the MxN region from the top, left quadrant of gp(x,y)

04-Dec-11 73
 An Example:
Steps for Filtering in the Frequency Domain

04-Dec-11 74
 Correspondence Between Filtering in the
Spatial and Frequency Domains (1)

Let H(u) denote the 1-D frequency domain Gaussian filter

- u 2 /2σ 2
H (u ) = Ae

The corresponding filter in the spatial domain

−2π 2σ 2 x 2
h( x) = 2πσ Ae

1. Both components are Gaussian and real

2. The functions behave reciprocally

04-Dec-11 75
 Correspondence Between Filtering in the
Spatial and Frequency Domains (2)
Let H (u ) denote the difference of Gaussian filter
- u 2 /2σ12 - u 2 /2σ 2 2
H (u ) = Ae − Be
with A ≥ B and σ 1 ≥ σ 2

The corresponding filter in the spatial domain

−2π 2σ12 x 2 −2π 2σ 22 x 2
h( x) = 2πσ 1 Ae − 2πσ 2 Ae

High-pass filter or low-pass filter ?

04-Dec-11 76
 Correspondence Between Filtering in the
Spatial and Frequency Domains (3)

04-Dec-11 77
 Correspondence Between Filtering in the
Spatial and Frequency Domains: Example

600x600

04-Dec-11 78
 Correspondence Between Filtering in the
Spatial and Frequency Domains: Example

04-Dec-11 79
 Generate H(u,v)

 f ( x, y ) 0 ≤ x ≤ 599 and 0 ≤ y ≤ 599

f p ( x, y ) = 
 0 600 ≤ x ≤ 602 or 600 ≤ y ≤ 602

 h ( x, y ) 0 ≤ x ≤ 2 and 0 ≤ y ≤ 2
h p ( x, y ) = 
0 3 ≤ x ≤ 602 or 3 ≤ y ≤ 602

Here P ≥ A(600) + C (3) − 1 = 602;

Q ≥ B(600) + D(3) − 1 = 602.

04-Dec-11 80
 Generate H(u,v)

1. Multiply hp ( x, y ) by (-1) x + y to center the frequency domain filter

2. Compute the forward DFT of the result in (1)

3. Set the real part of the resulting DFT to 0 to account for

parasitic real parts

4. Multiply the result by (-1)u + v , which is implicit when h( x, y )

was moved to the center of hp ( x, y ).

04-Dec-11 81
 Image Smoothing Using Filter Domain Filters:
ILPF

Ideal Lowpass Filters (ILPF)

1 if D(u , v) ≤ D0
H (u , v) = 
0 if D(u, v) > D0

D0 is a positive constant and D (u , v) is the distance between a point (u , v)

in the frequency domain and the center of the frequency rectangle
2 1/2
D(u , v) = (u − P / 2) + (v − Q / 2) 
2

04-Dec-11 82
 Image Smoothing Using Filter Domain Filters:
ILPF

04-Dec-11 83
 ILPF Filtering Example

04-Dec-11 84
 ILPF
Filtering
Example

04-Dec-11 85
 The Spatial Representation of ILPF

04-Dec-11 86
 Image Smoothing Using Filter Domain Filters:
BLPF
Butterworth Lowpass Filters (BLPF) of order n and
with cutoff frequency D0
1
H (u , v) =
1 + [ D(u, v) / D0 ]
2n

04-Dec-11 87
04-Dec-11 88
 The Spatial Representation of BLPF

04-Dec-11 89
 Image Smoothing Using Filter Domain Filters:
GLPF

Gaussian Lowpass Filters (GLPF) in two dimensions is given

− D 2 ( u , v )/2σ 2
H (u , v) = e

By letting σ = D0
− D 2 ( u , v )/2 D02
H (u , v) = e

04-Dec-11 90
 Image Smoothing Using Filter Domain Filters:
GLPF

04-Dec-11 91
04-Dec-11 92
 Examples of smoothing by GLPF (1)

04-Dec-11 93
 Examples of smoothing by GLPF (2)

04-Dec-11 94
 Examples of smoothing by GLPF (3)

04-Dec-11 95
 Image Sharpening Using Frequency Domain
Filters

A highpass filter is obtained from a given lowpass filter

using

H HP (u, v) = 1 − H LP (u, v)

A 2-D ideal highpass filter (IHPL) is defined as

0 if D(u , v) ≤ D0
H (u , v) = 
1 if D(u , v) > D0

04-Dec-11 96
 Image Sharpening Using Frequency Domain
Filters

A 2-D Butterworth highpass filter (BHPL) is defined as

1
H (u , v) =
1 + [ D0 / D(u , v) ]
2n

A 2-D Gaussian highpass filter (GHPL) is defined as

− D 2 ( u , v )/2 D02
H (u, v) = 1 − e

04-Dec-11 97
04-Dec-11 98
 The Spatial Representation of Highpass
Filters

04-Dec-11 99
 Filtering Results by IHPF

04-Dec-11 100
 Filtering Results by BHPF

04-Dec-11 101
 Filtering Results by GHPF

04-Dec-11 102
 Using Highpass Filtering and Threshold for
Image Enhancement

BHPF
(order 4 with a cutoff
frequency 50)

04-Dec-11 103
 The Laplacian in the Frequency Domain

H (u , v) = −4π 2 (u 2 + v 2 )

H (u , v) = −4π 2 (u − P / 2) 2 + (v − Q / 2) 2 ) 
= −4π 2 D 2 (u , v)
The Laplacian image
∇ 2 f ( x, y ) = ℑ−1 { H (u , v) F (u , v)}

Enhancement is obtained
g ( x, y ) = f ( x, y ) + c∇ 2 f ( x, y ) c = -1
04-Dec-11 104
 The Laplacian in the Frequency Domain

The enhanced image

−1
g ( x, y ) = ℑ {F (u, v) − H (u, v) F (u, v)}
= ℑ−1 {[1 − H (u, v) ] F (u , v)}

{ }
= ℑ−1 1 + 4π 2 D 2 (u , v)  F (u , v)

04-Dec-11 105
 The Laplacian in the Frequency Domain

04-Dec-11 106
 Unsharp Masking, Highboost Filtering and
High-Frequency-Emphasis Fitering

g mask ( x, y ) = f ( x, y ) − f LP ( x, y )

f LP ( x, y ) = ℑ−1 [ H LP (u, v) F (u, v) ]

Unsharp masking and highboost filtering

g ( x, y ) = f ( x, y ) + k * g mask ( x, y )

{ }
g ( x, y ) = ℑ−1 1 + k * [1 − H LP (u , v) ] F (u , v)
= ℑ−1 {[1 + k * H HP (u , v)] F (u , v)}

04-Dec-11 107
 Unsharp Masking, Highboost Filtering and
High-Frequency-Emphasis Fitering

−1
g ( x, y ) = ℑ {[ k
1 + k2 * H HP (u, v)] F (u, v)}
k1 ≥ 0 and k2 ≥ 0

04-Dec-11 108
 Gaussian Filter
D0=40

High-Frequency-Emphasis Filtering
Gaussian Filter
K1=0.5, k2=0.75

04-Dec-11 109
 Homomorphic Filtering

f ( x, y ) = i ( x, y ) r ( x, y )
ℑ [ f ( x, y ) ] ≠
= ℑ [i ( x, y ) ] ℑ [ r ( x, y ) ] ?

z ( x, y ) = ln f ( x, y ) = ln i ( x, y ) + ln r ( x, y )

ℑ{ z ( x, y )} = ℑ{ln f ( x, y )} = ℑ{ln i ( x, y )} + ℑ{ln r ( x, y )}

Z (u , v) = Fi (u, v) + Fr (u, v)

04-Dec-11 110
 Homomorphic Filtering

S (u , v) = H (u , v) Z (u, v)
= H (u, v) Fi (u, v) + H (u , v) Fr (u, v)
s ( x, y ) = ℑ−1 {S (u, v)}
= ℑ−1 { H (u, v) Fi (u , v) + H (u, v) Fr (u , v)}
= ℑ−1 { H (u, v) Fi (u , v)} + ℑ−1 { H (u, v) Fr (u, v)}
= i '( x, y ) + r '( x, y )

g ( x, y ) = e s ( x , y ) = ei '( x , y ) e r '( x , y ) = i0 ( x, y )r0 ( x, y )

04-Dec-11 111
 Homomorphic Filtering

The illumination component of an image generally is

characterized by slow spatial variations, while the
reflectance component tends to vary abruptly

These characteristics lead to associating the low

frequencies of the Fourier transform of the logarithm of an
image with illumination the high frequencies with
reflectance.
04-Dec-11 112
 Homomorphic Filtering


H (u , v) = (γ H − γ L ) 1 − e
− c  D 2 ( u , v )/ D02 
   +γ
  L

Attenuate the contribution

made by illumination and
amplify the contribution made
by reflectance

04-Dec-11 113
 γ L = 0.25
Homomorphic
γH = 2
Filtering
c =1
D0 = 80

04-Dec-11 114
 Homomorphic Filtering

04-Dec-11 115
 Selective Filtering

Non-Selective Filters:
operate over the entire frequency rectangle

Selective Filters
operate over some part, not entire frequency rectangle
• bandreject or bandpass: process specific bands
• notch filters: process small regions of the frequency
rectangle

04-Dec-11 116
 Selective Filtering:
Bandreject and Bandpass Filters

H BP (u, v) = 1 − H BR (u , v)

04-Dec-11 117
 Selective Filtering:
Bandreject and Bandpass Filters

04-Dec-11 118
 Selective Filtering:
Notch Filters
Zero-phase-shift filters must be symmetric about the origin.
A notch with center at (u0, v0) must have a corresponding
notch at location (-u0,-v0).

Notch reject filters are constructed as products of highpass

filters whose centers have been translated to the centers of
the notches.
Q
H NR (u, v) = ∏ H k (u, v) H − k (u, v)
k =1

where H k (u , v) and H - k (u , v) are highpass filters whose centers are

at (uk , vk ) and (-uk , -vk ), respectively.
04-Dec-11 119
 Selective Filtering:
Notch Filters
Q
H NR (u , v) = ∏ H k (u , v) H − k (u , v)
k =1

where H k (u , v) and H - k (u, v) are highpass filters whose centers are

at (uk , vk ) and (-uk , -vk ), respectively.

A Butterworth notch reject filter of order n

3  1  1 
H NR (u , v) = ∏  2n   2n 
k =1 1 + [ D0 k / Dk (u , v ) ]  1 + [ D0 k / D− k (u , v ) ] 
  
2 1/ 2
Dk (u, v) = (u − M / 2 − uk ) + (v − N / 2 − vk ) 
2

2 1/ 2
D− k (u, v) = (u − M / 2 + uk ) + (v − N / 2 + vk ) 
2

04-Dec-11 120
Examples:
Notch
Filters (1)

A Butterworth notch
reject filter D0 =3
and n=4 for all
notch pairs

04-Dec-11 121
 Examples:
Notch Filters
(2)

04-Dec-11 122
04-Dec-11 123