Image Analysis and Processing

Image Enhancements
in the Frequency Domain
Laurent Najman
[email protected]

ESIEE Paris
Universite Paris-Est, Laboratoire dInformatique Gaspard-Monge, Equipe
A3SI

image processing, transforms p. 1/46

 Book
Chapter 4 (pages 147219)

Digital Image Processing, Second Edition

authors: Rafael C. Gonzalez and Richard E. Woods
editor: Prentice Hall

p.3/44
Frequency Domain Filtering Operation
Frequency domain: space defined by values of the Fourier transform and its
frequency variables (u, v).

Relation between Fourier Domain and image:

u = v = 0 corresponds to the gray-level average
Low frequencies: images component with smooth gray-level variation (e.g.
areas with low variance)

High frequencies: quick gray-level change (e.g. edges, noise)

Source: http://www.imageprocessingbook.com p.4/44

 Filtering in the Frequency Domain
Let H(u, v) a filter, also called filter transfer function.
Filter: suppress certain frequencies while leaving
others unchanged
G(u, v) = H(u, v) F (u, v)

H(u, v) in image processing:

In general H(u, v) is real: zero-phase-shift filter
H multiply real and imaginary parts of F
I(u,v)
(u, v) = tan1 R(u,v) does not change if H is real

p.5/44
 General Steps for Filtering
1. Multiply input image by (1)x+y (centering)
2. Compute F (u, v) (DFT)
3. Multiply F (u, v) by H(u, v) (filtering)
4. Compute inverse DFT of H(u, v) F (u, v)
5. Obtain the real part of the result
6. Mutliply by (1)x+y (decentering)

p.6/44
From Spatial to Frequency Domain

f (x, y) h(x, y) F (u, v) H(u, v)

(x, y) h(x, y) F [(u, v)] H(u, v)
h(x, y) H(u, v)

Multiplication in the frequency domain is a convolution

in the spatial domain.
Given h(x, y), we can obtain H(u, v) by taking the
inverse Fourier transform.

p.7/44
 Padding (1)
convolution:
M 1
1 X
f (x) h(x) = f (m) h(x m)
M
m=0

periodicity:
F (u, v) = F (u + M, v) = F (u, v + N ) = F (u + M, v + N )

p.8/44
Padding (2)

Source: http://www.imageprocessingbook.com p.9/44

 Padding (3)
Let P an identical period for f and g :
(
f (x) 0 x A 1
fe (x) =
0 AxP

(
g(x) 0 x B 1
ge (x) =
0 BxP
If P < A + B 1, the two signal will overlap:
wraparound error.
If P > A + B 1, the periods will be separated.
If P = A + B 1, the periods will be adjacent.

We can avoid wraparound error using P A + B 1.

In general, we use P = A + B 1.
p.10/44
Padding (4)

Source: http://www.imageprocessingbook.com p.11/44

Padding (5)

Source: http://www.imageprocessingbook.com p.12/44

Padding: Example

Source: http://www.imageprocessingbook.com p.13/44

Spatial Representation of a Filter
1. Multiply filter H(u, v) by (1)u+v (centering)
2. Compute the inverse DFT
3. Multiply the real part of the inverse DFT by (1)x+y

Source: http://www.imageprocessingbook.com p.14/44

A Simple Filter: Notch Filter (1)
We wish to force the average value of an image to
zero:
F (0, 0) is the average value of the image
if size of the image is M N then the centered
value of the Fourier transform is the average value
(M 2 , 2)
N

(
0 if (u, v) = ( M
2 2)
, N
H(u, v) =
1 otherwise.

p.15/44
A Simple Filter: Notch Filter (2)

notch filter (constant function with a hole at the origin)

(
0 if (u, v) = ( M
2 , N
2 )
H(u, v) =
1 otherwise.

Source: http://www.imageprocessingbook.com p.16/44

Lowpass and Highpass Filter

Source: http://www.imageprocessingbook.com p.17/44

Smoothing Frequency-Domain Filters

G(u, v) = H(u, v) F (u, v)

Smoothing: attenuating specified range of

high-frequency components
Three types of lowpass filter:
ideal (very sharp)
Butterworth (tunable)
Gaussian (very smooth)

p.18/44
 Ideal Lowpass Filters (1)
(
1 if D(u, v) D0
H(u, v) =
0 if D(u, v) > D0
D0 : nonnegative quantity
D(u, v): distance from a point (u, v) to the origin
origin: ( M
2 2 ) (centered)
, N

Then,
r
M 2 N 2
D(u, v) = (u ) + (v )
2 2

p.19/44
 Ideal Lowpass Filters (2)

ideal filter (ILPF): all frequencies inside a circle of

radius D0 are passed with no attenuation
In our case:
zero-phase-shift filter
radially symmetric
Transition: cutoff frequency
Ringing behavior of the filter Source: http://www.imageprocessingbook.com p.20/44
Ideal Lowpass Filters: Example

10 (13%) 30 (6.9%)

60 (6.9%) 160 (4.3%)

460 (0.8%)
160 (2.2%)

Source: http://www.imageprocessingbook.com p.21/44

 Cutoff Frequency
Compute circles that enclose specific amount of total
image power PT

M
X 1 N
X 1
PT = P (u, v)
u=0 v=0

P (u, v) = |F (u, v)|2 = R2 (u, v) + I 2 (u, v)

100 X X
= P (u, v)
PT u v

Source: http://www.imageprocessingbook.com p.22/44

Ringing Effect of ILPF

Source: http://www.imageprocessingbook.com p.23/44

 Butterworth Lowpass Filters
BLPF of order n, with a cutoff frequency distance D0 is
defined as
1
H(u, v) =
1 + [ D(u,v)
D0 ] 2n

r
M 2 N 2
D(u, v) = (u ) + (v )
2 2

no clear cutoff between passed and filtered

frequencies

p.24/44
Butterworth Lowpass Filters (2)

Source: http://www.imageprocessingbook.com p.25/44

Butterworth Lowpass Filters: Example
n = 2, radii=5, 15, 30, 80, 230

Source: http://www.imageprocessingbook.com p.26/44

Ringing Effect of BLPF

Source: http://www.imageprocessingbook.com p.27/44

 Gaussian Lowpass Filters (1)
The form of a gaussian lowpass filter GLPF in 2D is:

D 2 (u,v)
H(u, v) = e 2 2

r
M 2 N 2
D(u, v) = (u ) + (v )
2 2
The inverse Fourier transform of a GLPF is also a
Gaussian
A spatial Gaussian filter will have no ringing

p.28/44
 Gaussian Lowpass Filters (2)
: mesure of the spread of the Gaussian curve
Let = D0 , then:
D 2 (u,v)
2
H(u, v) = e 2D0

Source: http://www.imageprocessingbook.com p.29/44

Gaussian Lowpass Filters: Example (1)
n = 2, radii=5, 15, 30, 80, 230

Source: http://www.imageprocessingbook.com p.30/44

Gaussian Lowpass Filters: Example (2)

Source: http://www.imageprocessingbook.com p.31/44

 Highpass Filters (1)
Highpass filter: image sharpening (low-frequency
attenuation)
In our case:
zero-phase-shift filter
radially symmetric

Hhp (u, v) = 1 Hlp (u, v)

with:
Hlp (u, v): transfer function of the corresponding
lowpass filter
Hhp (u, v): transfer function of the corresponding
highpass filter

p.32/44
Highpass Filters (2)

Source: http://www.imageprocessingbook.com p.33/44

 Ideal Highpass Filters (1)
(
0 if D(u, v) D0
H(u, v) =
1 if D(u, v) > D0
D0 : nonnegative quantity
D(u, v): distance from a point (u, v) to the origin
origin: ( M
2 2 ) (centered)
, N

Then,
r
M 2 N 2
D(u, v) = (u ) + (v )
2 2

p.34/44
Ideal Lowpass Filters: example

Source: http://www.imageprocessingbook.com p.35/44

Butterworth Highpass Filters
BLPF of order n, with a cutoff frequency distance D0 is
defined as
1
H(u, v) = D0 2n
1 + [ D(u,v) ]
r
M 2 N 2
D(u, v) = (u ) + (v )
2 2

no clear cutoff between passed and filtered

frequencies

p.36/44
Butterworth Highpass Filters: Example

Source: http://www.imageprocessingbook.com p.37/44

 Gaussian Highpass Filters
The form of a gaussian lowpass filter GLPF in 2D is:

D 2 (u,v)
2
H(u, v) = 1 e 2D0

r
M 2 N 2
D(u, v) = (u ) + (v )
2 2
The inverse Fourier transform of a GLPF is also a
Gaussian
A spatial Gaussian filter will have no ringing

p.38/44
Gaussian Highpass Filters: Example

Source: http://www.imageprocessingbook.com p.39/44

Laplacian: Spatial Domain (1)
2f 2f
2 f = x2 + y 2
2f
x2 = f (x + 1, y) + f (x 1, y) 2f (x, y)
2f
y 2 = f (x, y + 1) + f (x, y 1) 2f (x, y)
2 f = f (x + 1, y) + f (x 1, y) + f (x, y + 1)
+ f (x, y 1) 4f (x, y)

Source: http://www.imageprocessingbook.com p.40/44

 Laplacian: Spatial Domain (2)
(
f (x, y) 2 f (x, y) if the center coefficient is negative
g(x, y) =
f (x, y) + 2 f (x, y) if the center coefficient is positive

Source: http://www.imageprocessingbook.com p.41/44

 Laplacian: Frequency Domain (1)
 n
d f (x)

F = (ju)n F (u)
dxn
Then,
2 2
h i
f (x,y) f (x,y)
F x2 + y 2 = (ju)2 F (u, v) + (jv)2 F (u, v)
= (u2 + v 2 ) F (u, v)

Finally,
F [2 f (x, y)] = (u2 + v 2 )F (u, v)
The Laplacian filter in the frequency domain is:

H(u, v) = (u2 + v 2 )

p.42/44
Laplacian: Frequency Domain (2)

p.43/44
Laplacian: Example

p.44/44