Image Enhancement in Frequency

Domain
 Module - 2 : Image Transforms

 Typically interested in Fourier Transform (FT)

2
 Let us try to attach
“meaning” to the frequency
domain with respect to Image
Processing

3
Spatial Domain  Frequency Domain
 In image processing:
Instead of time domain: spatial domain
(normal image space)

Frequency domain: space in which each

image value at image position F represents
the amount that the intensity values in image
I vary over a specific distance related to F

4
What is Spatial Frequency?

 Spatial frequency of an image refers to

the rate at which the pixel intensities
change

6
Revision of FT
 FT decomposes a signal into a set of
sines/cosines of different characteristics like
frequency and phase

 FT is perfectly reversible

 If Fourier coefficients are multiplied by suitably

chosen weighing function
 Certain frequency components can be accentuated
while others are attenuated
Frequency Domain Filtering
 Selective enhancement or suppression of
frequency components is termed as Fourier
Filtering or Frequency Domain Filtering

 Spatial representation  adjacency

relationship between the pixels

 Frequency representation  clusters the

image data according to their frequency
distribution
Frequency Domain and Images
 General comments about relationship between
the frequency components of FT and spatial
characteristics of an image
The slowest varying frequency component ( u = v =
0) corresponds to the average gray level of an
image (origin)
Moving away from the origin  low frequencies
correspond to slowly varying components of an
image
Moving further away from the origin  higher
frequencies i.e. faster gray level changes (edges of
objects and other abrupt changes such as noise)

9
Frequency Domain and Images

v
(-N/2, -N/2) (-N/2, N/2)
high high

low
u
high high

(N/2, -N/2) (N/2, N/2)

10
Frequency Domain and Images

protrusions
Higher frequencies
correspond to the
faster varying intensity
level changes in the
image. The edges of
objects or the other
components
characterized by the Notice the ±45°
abrupt changes in the components and the
intensity level vertical component
which is slightly off-
corresponds to higher
axis to the left! It
frequencies corresponds to
the protrusion caused
by thermal failure
above.
11
Image Filtering in Frequency
Domain

= G(u,v)

12
A simple Demonstration – Notch filter
 Multiply all values of F(u,v) by the filter function (notch filter):

 0 if (u , v)  ( M / 2, N / 2)
H (u , v)  
1 otherwise.
 Set F(0,0) to zero (force the average value of an image to
zero) and leave all frequency components of the Fourier
transform untouched.
Low-pass Filtering in Frequency
Domain

 Standard forms of low-pass filters (LPF)

Ideal low-pass filter

Butterworth low-pass filter

Gaussian low-pass filter

14
Ideal Low-pass Filter
 The simplest low-pass filter is a filter that “cuts off” all
high-frequency components of the Fourier transform that
are at a distance greater than a specified distance D0
from the origin of the transform (D0=cut-off frequency)
 Transfer function of an ideal low-pass filter
1 if D(u , v)  D0
H (u , v)  
0 if D(u , v)  D0

where D(u,v) : the distance from point

(u,v) to the center of the frequency rectangle

 
1
D(u , v)  (u  M / 2) 2  (v  N / 2) 2 2

15
Ideal Low-pass Filter

16
Example Image and its FT

17
 Ideal LPF
D0 = 5

Note :
Narrower the D0 = 15 D0 = 30
filter in the
frequency
domain, more
severe are the D0 = 80 D0 = 230
blurring and
ringing!

18
Butterworth Low-pass Filter (BLPF)
 Transfer function is given by
1
H (u , v)  2n
 D(u , v) 
1  
 D 0 

where n = filter order, D0 = cut-off frequency

 
1
D(u , v)  (u  M / 2) 2  (v  N / 2) 2 2

20
Butterworth Low-pass Filter (BLPF)
1
H (u , v )  2n
 D (u , v ) 
1  
 D 0 

21
% This code is used to perform Butterworth low-pass filter
clc
im = imread('cameraman.tif');
D0=20; %cut-off frequency
n=1; %order
[M N] = size(im);
% find the centre of the image
cx = round(N/2);
cy = round (M/2);
imf = fftshift(fft2(im));
H = zeros (M, N);
for u = 1:N
for v = 1:M
d = sqrt((u - cx).^2 + (v - cy).^2);
H (u,v) = 1/(1+(d/D0).^(2*n));
end
end
outf = imf .* H;
out = abs (ifft2(outf));
subplot(2,1,1); imshow(im); title ('Original Image');
subplot(2,1,2); imshow(uint8(out)), title('Lowpass Filtered Image (BLPF)');

22
Butterworth Low-pass Filter (BLPF)
The smaller
the value of
D0, the greater
the number of
image
components
eliminated by
the filter

23
BLPF
Order, n=2

Cut-off frequency,
D0=5,15,30,80,and 230

24
BLPF

 No ringing for n=1, imperceptible ringing for n=2, ringing

increases for higher orders (getting closer to Ideal LPF)
25
Gaussian Low-pass Filter (GLPF)
 Transfer Function

2 2
D
H (u , v)  e ( u ,v ) / 2 D0

 
1
D(u , v)  (u  M / 2) 2  (v  N / 2) 2 2

26
Gaussian Low-pass Filter (GLPF)

27
% This code is used to perform Gaussian low-pass filter
clc
im = imread('cameraman.tif');
D0=10; %cut-off frequency
[M N] = size(im);
% find the centre of the image
cx = round(N/2);
cy = round (M/2);
H = zeros (M, N);
imf = fftshift(fft2(im));
for u = 1:N
for v = 1:M
d = (u - cx).^2 + (v - cy).^2;
H (u,v) = exp(-d/D0.^2);
end
end
outf = imf .* H;
out = abs (ifft2(outf));
subplot(2,1,1); imshow(im); title ('Original Image');
subplot(2,1,2); imshow(uint8(out)), title('Low-pass Filtered Image (GLPF)');

28
GLPF

Cut-off frequency,
D0=5,15,30,80,and 230

 No ringing

29
30
Additional Examples of GLPF

31
Additional Examples of GLPF

“Cosmetic” processing
LPF – to produce a
smoother, softer-looking
results from sharp originals

LPF is used in printing, e.g. to smooth fine skin lines in faces. 32

 High-pass Filtering in Frequency
Domain
H hp (u , v)  1  H lp (u , v)
Ideal high-pass filter

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

Butterworth high-pass filter

1
H (u , v) 
1  D0 / D(u , v)
2n

Gaussian high-pass filter

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

33
Ideal High-pass Filter
0 if D(u , v)  D0
H (u , v)  
1 if D(u , v)  D0

Ideal high-pass filters enhance edges but suffer from ringing artifacts,
just like Ideal LPF. 34
% This code is used to perform Butterworth high-pass filter
clc
im = imread('cameraman.tif');
D0=40; %cut-off frequency
n=1; %order
[M N] = size(im);
% find the centre of the image
cx = round(N/2);
cy = round (M/2);
imf = fftshift(fft2(im));
H = zeros (M, N);
for u = 1:N
for v = 1:M
d = sqrt((u - cx).^2 + (v - cy).^2);
H (u,v) = 1/(1+(D0/d).^(2*n));
end
end
outf = imf .* H;
out = abs (ifft2(outf));
subplot(2,1,1); imshow(im); title ('Original Image');
subplot(2,1,2); imshow(uint8(out)), title(‘High-pass Filtered Image (BHPF)');

35
Butterworth High-pass Filter
1
H (u , v) 
1  D0 / D(u , v)
2n

Improved enhanced images with BHPFs

36
Gaussian High-pass Filter
 D 2 ( u ,v ) / 2 D02
H (u , v)  1  e

Even smoother results with GHPFs 37

Exercise
 Compare
Ideal, Butterworth and Gaussian LPF
Ideal, Butterworth and Gaussian HPF

38
 Comments on LPFs
 ILPF
 The spatial filter
h (x, y) has two
major
characteristics
○ Dominant
component at
the origin
○ Concentric
circular
components
about center
component

Center component is responsible for blurring

Circular components are responsible for the ringing artifacts
39
Comments on LPFs
 BLPF
 The BLPF with order of 1 does not have any ringing
artifact.
 BLPF with orders 2 or more shows increasing ringing
effects as the order increases.
Comments on HPFs
 IHPF
The ringing artifacts occur at low cutoff
frequencies
 BHPF
Smoother results are obtained in BHPF
when compared IHPF. There is almost no
ringing artifacts
 GHPF
Smoother results are obtained in GHPF
when compared BHPF. There is absolutely
no ringing artifacts
 Comparison of Ideal, Butterworth and
Gaussian High pass Filters (D0 = 30)

43
The Laplacian in Frequency
Domain
 Transfer function

H (u, v)   D (u, v)2

where

 
1
D(u , v)  (u  M / 2)  (v  N / 2)
2 2 2

44
The Laplacian in Frequency
Domain
 Laplacian-filtered image in spatial
domain

1
 f ( x, y )   {H (u, v) F (u, v)}
2

Need to revisit Laplacian in Spatial Domain ?

45
 Laplacian In The Frequency
Domain

2-D image of Laplacian

frequency domain

in the frequency
Laplacian in the

domain
frequency domain
Laplacian in the
Inverse DFT of

Zoomed
section of
the image on
the left
compared to
spatial filter
The Laplacian in Frequency
Domain

47
Unsharp Masking

f hp ( x, y )  f ( x, y )  f lp ( x, y )
With

1
f lp ( x, y )   [ H lp (u , v) F (u , v)]
Low pass filter FT of f ( x, y)
High-boost Filtering
 Generalization

f hb ( x, y )  Af ( x, y )  f lp ( x, y )
Where A ≥ 1

f hb ( x, y )  ( A  1) f ( x, y )  f hp ( x, y )
Unsharp masking and High-boost
filtering in Frequency Domain
Fhp (u , v)  F (u , v)  Flp (u , v)
but
Flp (u , v)  H lp (u , v) F (u , v)

 Direct Implementation by using composite filter

H hp (u , v )  1  H lp (u , v )
And
H hb (u , v )  ( A  1)  H hp (u , v )
50
Example – High-boost
Filtering

51
High-frequency Emphasis

 Objective
To accentuate the contribution to
enhancement made by high-frequency
components of an image

 Multiply a high-pass filter function by a

constant and add an offset

52
High-Frequency Emphasis
 Transfer function

H hfe  a  bH hp (u , v)
where a  0 and b  a
 Typical values of a  0.25 to 0.5
 Typical values of b  1.5 to 2.0

 HFE reduces to HBF when a = (A – 1) and b = 1

 When b > 1, high frequencies are emphasized
53
 Example – High-Frequency Emphasis
Filtering

Highpass filtering result

Original image

a=0.5, b= 2.0
emphasis result
High frequency

After histogram
equalisation
Highpass Filtering Example
Highpass Filtering Example
Highpass Filtering Example
Highpass Filtering Example
Homomorphic Filtering
 Many times, we want to remove shading
effects from an image (i.e., due to uneven
illumination)
Enhance high frequencies
Attenuate low frequencies but preserve fine detail.

59
Homomorphic Filtering
 Consider the following model of image formation:


( x, y )  i ( x, y ) r ( x, y )
fIn general, I (x,y): illumination
r(x, y): reflection

The illumination component i (x, y) varies slowly and

affects low frequencies mostly.
The reflection component r (x, y) varies faster and
affects high frequencies mostly.

IDEA: separate low frequencies due to i (x, y)

from high frequencies due to r (x,y)
Homomorphic Filtering

61
Homomorphic Filtering
 c ( D 2 ( u ,v ) / D02 )
H (u , v)  ( H   L )[1  e ] L

L <1 H >1

Attenuate the contribution

Attenuate the
made by illumination
contribution made by (low
frequencies)
illumination and amplify the
and amplify
contribution
the contributionmade
madeby
by reflectance
reflectance (high frequencies)
Example - Homomorphic
Filtering
L =0.5 H =2.0

63
Example - Homomorphic
Filtering

 L  0.25
H  2
c 1
D0  80
A Final Comment
Similar jobs can be done in the spatial and
frequency domains
Filtering in the spatial domain can be
easier to understand
Filtering in the frequency domain can be
much faster – especially for large images
Why Frequency Domain?
-One Reason

66