03 Image Enhancement in The Frequency Domain (Chapter 04)

CSE483 Computer Vision
Image Enhancement in the Frequency

Domain
Very incredibly annoying lecture, would be best to
study it in further depth from the book - Chapter 4.
Prof. Mahmoud Khalil

Spring 2022
1
Why do we bother converting to Frequency Domain? This has an overhead of its own of course...
Well, one reason among many is for example JPEG compression which uses frequency domain.
Image Enhancement in the Frequency Domain
2
Very hard to point out the

noise in terms of spatial
domain, but once you go
frequency, it's very easy to
notice the strong sinusoidal
noise pattern, which means
over the frequency domain it
will have a strong feedback
and thus be easily removable.
3
4
The 1D Continuous Fourier Transform
The continuous Fourier transform
The inverse continuous Fourier transform
5
The 1D Discrete Fourier Transform (DFT)
The 1D discrete Fourier transform
The inverse 1D discrete Fourier transform
6
Each element from both sides has

an effect on all the elements7 of
the other side.
General rule of thumb: If the function width is large, its transform is small. The middle
position would be the Gaussian wave where its input is its output too. Another property
is duality, the input and output are linked in either transformation direction.
8
9
Makes sense that output be an

impulse there. It's already a basis
function (has cosine and sine), so
obviously its value is just one value
in the frequency domain, at the
frequency of the cosine and sine.
10
11
13
The 2D discrete Fourier transform
u: horizontal axis
v: vertical axis
The 2D inverse discrete Fourier transform
14
Both are sized M*N
15
16
Relative to the image above,

this one is wider, hence its
transform will be less wide.
17
In the previous slide, when the wave was wide, its transform was thin, and vice versa...
Here, you have a 2D rectangle in the spatial domain. Its transform will thus also be a rectangle. As for the
dimensions, apply the same thinking: The wider dimension in spatial domain will be the thinner one in freq.

Notice that the width of the horizontal
blocks is smaller than the vertical ones;
while the original image was wider
horizontally than vertically...
18
Note how the "y" and "v" axes are in the same direction (even though they are
Notice how the origin is
horizontal while they should be vertical but whatever) and the "x" and "u" too. centered in freq. domain.
Translation Eqn in the end.
Fourier Transform ‐ Examples
2D Image 2D Image - Rotated
Fourier Spectrum Fourier Spectrum
19
Image Frequency
Domain Domain
23
Image
Fourier spectrum
28
Notice how everything above has a feedback

here, perpendicular to it. The most notable:
(1) El beta3a ely beyed8at biha di
(2) The thing that is being pressed
29
(3) The edges of the squared things under
The Fourier Image Plotting the log helps us see all the values better, cause
if naturally the difference between the very low and
very high values is big, the log reduces that difference.
This difference is usually large for real pictures naturally.
Fourier spectrum |F(u,v)| Fourier spectrum log(1 + |F(u,v)|)

Image f
30
Frequency Bands
Image Fourier Spectrum
Percentage of image power enclosed in circles (small to large) :
90%, 95%, 98%, 99%, 99.5%, 99.9%
31
32
This will have a global effect since as we

mentioned earlier each component has an
effect on all the other components, and
also as observed the central components
are usually the strongest... In fact, F(0, 0) is
the DC component of the transform! So,
setting it to 0 means everything will move
down greatly, some going negative,
meaning that we'll need to add some value 33
to the result image to be able to see it
(obv. though if we re-add the DC value it'd
be back to how it was lol)
Observe how LPF looks like smoothing,
and HPF looks like sharpening... 34
35
Mexican Hat
These curves at the bottom, the

fourier transforms of the HPF and
LPF above, make sense to be the
36
continuous versions of their
corresponding masks (specifically
the ones pointed at), so yeah...
LPF = Smoothen, HPF = Sharpen
Laplacian Filter in the Frequency Domain
37
Problem 4.27 The fact that he said "averages" means
he didn't need to tell you it'll be lowpass
Suppose that you form a lowpass spatial filter that averages
the four immediate neighbors of a point (x, y) but excludes
the point itself.
(a) Find the equivalent filter H(u, v) in the frequency

domain.
(b) Show that your result is a lowpass filter.

It has cosines, cosines look like a lowpass hull. Done.
38
Ideal Lowpass Filters
The hyper-parameter,
the cut-off frequency.
39
40
41
These filters perform compression, but the result should still look proper to be
acceptable... The ringing we saw previously was a violation of the rule that we
said earlier for averaging filters, where if an area is smooth, it must stay so.
Butterworth Lowpass Filters

The higher the n, the
more ideal it gets, the
more ringing that occurs
So now there are two hyper-parametrs, the

cut-off frequency D0 and the filter order n.
43
44
45
Gaussian Lowpass Filters
46
47
48
Doesn't seem like the output is different. The difference is
actually in performance. Going to frequency domain is an
overhead, but is the algorithm even more demanding?
Highpass Filters
49
Highpass Filters
50
Ideal Highpass Filters
51
Butterworth Highpass Filters
52
Gaussian Highpass Filters
53
Noise Removal
Noisy image
But how to know which

frequencies are noise
and which are original
image material?@
Fourier Spectrum Noise-cleaned image 54
Noise Removal
Noisy image Fourier Spectrum Noise-cleaned image
55

03 Image Enhancement in The Frequency Domain (Chapter 04)

Uploaded by

Copyright:

Available Formats

03 Image Enhancement in The Frequency Domain (Chapter 04)

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

03 Image Enhancement in The Frequency Domain (Chapter 04)

Uploaded by

Copyright:

Available Formats

CSE483 Computer Vision

Image Enhancement in the Frequency

Prof. Mahmoud Khalil

Image Enhancement in the Frequency Domain

Very hard to point out the

The inverse continuous Fourier transform

The inverse 1D discrete Fourier transform

Each element from both sides has

The 1D Discrete Fourier Transform (DFT)

Makes sense that output be an

The 2D inverse discrete Fourier transform

Both are sized M*N

Relative to the image above,

Image Enhancement in the Frequency Domain

Fourier Spectrum Fourier Spectrum

Notice how everything above has a feedback

Fourier spectrum |F(u,v)| Fourier spectrum log(1 + |F(u,v)|)

Image Fourier Spectrum

Percentage of image power enclosed in circles (small to large) :

90%, 95%, 98%, 99%, 99.5%, 99.9%

This will have a global effect since as we

These curves at the bottom, the

(a) Find the equivalent filter H(u, v) in the frequency

(b) Show that your result is a lowpass filter.

Butterworth Lowpass Filters

So now there are two hyper-parametrs, the

But how to know which

Noisy image Fourier Spectrum Noise-cleaned image

You might also like