Week4 - Frequency Domain Operation

EIE4512 - Digital Image Processing
Frequency Domain Operations
Zhen Li
lizhen@cuhk.edu.cn
School of Science and Engineering
The Chinese University of Hong Kong, Shen Zhen
February 19-21, 2019
Basics of Frequency
Domain
1
Outline
• Preliminary:
• Complex Numbers
• Fourier Series & Fourier Transform
• Impulses
• Convolution
• Sampling
• Sampled Function
• Sampling Theorem
2
Revisit of Complex Numbers
• Definition
• ! = # + % '
• Conjugate: ! ∗ = # − % '
• Polar Coordinates '

! = # + %'
• ! = ! (cos . + % sin .)
|!|
• ! = #2 + '2
• tan . = '/# .
#
• Euler’s Formula
• 6 78 = cos . +jsin .
• ! = |!|6 78
3
Fourier Series
• Any periodic function !(#) with period % can be
represented as a Fourier Series
1
X
j 2⇡n
f (t) = cn e T
n= 1
• With
Z T /2
1 j 2⇡n
cn = f (t)e T t dt
T T /2
Fourier Series for real value function and conplex value function 4
https://www.youtube.com/watch?v=kP02nBNtjrU
Chapter 4
Filtering in the Frequency Domain
Unit Impulse
4.2.3 Impulses and their Shifting Property
Unit impulse of a continuous variable t located at t=0 :
and
Sifting property :
(if f(t) continuous at t=0 )
© 1992–2008 R. C. Gonzalez & R. E. Woods

5
Digital Image Processing
Discrete Unit Impulse

T. Peynot
Chapter 4
Unit discrete impulse located at x=0 :
Sifting property :
6
T. Peynot
Chapter 4
Impulse Train Filtering in the Frequency Domain
Impulse train : sum of infinitely many periodic impulses ΔT units apart:
7
Chapter 4
Fourier Transform (FT) of Filtering in the Frequency Domain
Continuous Functions
4.2.4 The Fourier Transform of Functions of One Continuous Variable
Fourier Transform of a continuous function f(t) of a continuous variable t :
Inverse Fourier Transform:
8
T. Peynot
FT Example
Chapter 4
Example :

9
T. Peynot
FT of Unit Impulse Chapter 4

Example 2 : Fourier Transform of a unit impulse located at the origin:
Fourier Transform of a unit impulse located at t = t0 :
10
T. Peynot
FT of Impulse Train Chapter 4

Example 2 : Fourier Transform S(µ) of an impulse train with period ΔT :

Periodic function with period ΔT =>
Where :
=>
Also an impulse train, with period 1/ΔT
© 1992–2008 R. C. Gonzalez & R. E. Woods 11

T. Peynot
Chapter 4
Continuous-Time Convolution Filtering in the Frequency Domain
4.2.5 Convolution
Convolution of functions f(t) and h(t), of one continuous variable t :
Fourier Transform pairs:
Spatial Domain Conv <=> Frequency Domain Multiplication
Spatial Domain Multiplication <=> Frequency Domain Conv

12
Commonly used continuous Fourier transform
T. Peynot
Chapter 4
Chapter 4
Sampling
4.3 Sampling and the Filtering in the Frequency
Fourier Transform Domain
of Sampled Functions
4.3.1 Sampling
4.3 Sampling and the Fourier Transform of Sampled Functions
Model of the sampled function:
4.3.1 Sampling of f(t) by a sampling
multiplication
function equal to a train of impulses ΔT
units apart
Model of the sampled function:
multiplication of f(t) by a sampling
function equal to a train of impulses ΔT
units apart
1992–2008 R. C. Gonzalez & R. E. Woods

13
T. Peynot
FT of Sampled Function Chapter 4

h i
F̃ (µ) = F T f (t)4.3.3 The
˜ Sampling Theorem
= F T [f (t) · s T (t)]
= F (µ) ? S(µ)
Z 1
= F (⌧ )S(µ ⌧ )d⌧
1
Z ⇣ n ⌘
1 1
X
1
= F (⌧ ) µ ⌧ d⌧
T 1 n= 1
T
Z ⇣ n ⌘
1
X 1
1
= F (⌧ ) µ ⌧ d⌧
T n= 1 1 T
1 X1 ⇣ n ⌘
= F µ
T n= 1
T
Superposition of infinitely many

copies of ! " with an interval 1/∆&
14
T. Peynot
Sampling Theorem Chapter 4

A function f(t) whose Fourier transform is zero

for values of frequencies outside a finite
interval (band) [- µmax, µmax] about the origin is
called a band-limited function
Sufficient separation guaranteed if:
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
NB: Sampling at:

Nyquist rate
15
T. Peynot
Recovery with Low-pass Filter Chapter 4

=> Recover f(t) using the inverse FT:
Example: ideal lowpass filter
16
T. Peynot
Aliasing
Chapter 4
4.3.4 Aliasing
Effect of under-sampling a function

Transform corrupted by frequencies
from adjacent periods
NB: the effect of aliasing can be reduced

by smoothing the input function to
attenuate its higher frequencies: anti-
aliasing.
Has to be done before the sampling
17
Chapter 4
Aliasing (Illustration)
4.3.4 Aliasing
Illustration: sampling a sine wave of period 2s
18
Function Reconstruction
h i
1
f (t) = F T F̃ (µ)H(µ)
1
X 
1
= f (n T ) · sinc (t n T)
n= 1
T
https://blog.csdn.net/u010592995/article/details/73555425
https://blog.csdn.net/liyuanbhu/article/details/42061089
19
Discrete Fourier Transform
FT of a sampled function
Z 1
F̃ (µ) = f˜(t)e j2⇡µt
dt
1
Z 1 1
X
j2⇡µt
= f (t) (t n T )e dt
1 n= 1
1 Z 1
X
j2⇡µt
= f (t) (t n T )e dt
n= 1 1
X1
j2⇡µn T
= fn e
n= 1
Consider a finite sequence of samples
⇣ m ⌘
M
X1
j2⇡unm/M
DFT: Fm = F̃ = fne
M T n=0
20
DFT and IDFT
Discrete Fourier Transform (DFT)
M
X1
j2⇡mn/M
Fm = fn e
n=0
Inverse Discrete Fourier Transform (IDFT)
M
X1
1
fn = Fm ej2⇡mn/M
M m=0
21
Relation between Sampling &
Frequency Interval
• Consider ! samples of ", with interval ∆$
• The duration of the record is

• T = ! ( ∆$
• The entire frequency range is

)
• ∆*
• The spacing in the frequency domain

) )
• ∆+ = , ∆* = *
22
Chapter 4
Circular Convolution
4.4 The Discrete Fourier Transform (DFT) of One Variable
4.4.1 Obtaining the DFT from the Continuous Transform of a Sampled Function
Both forward and inverse discrete transforms are infinitely periodic, with period M
Where k is an integer
Discrete equivalent of the convolution:
23
Chapter 4
Extension to Function of Two Variables

4.5 Extension to Function of Two Variables
4.5.1 The 2-D impulse and its Sifting Property
Impulse of two continuous variables t and z :

T. Peynot
Chapter 4
Impulse of two continuous variables t and z :
4.5.1 The 2-D impulse and its Sifting Property
Sifting property:
24
Extension
4.5.1 The 2-D of Discrete
impulse and its SiftingImpulse
Property
2-D discrete impulse:
Sifting property:

Chapter 4
Extension of Fourier
4.5 Extension to Function of Two VariablesTransform
4.5.2 The 2-D Continuous Fourier Transform Pair
f(t,z) continuous function of two continuous variables t and z, the 2-D continuous
Fourier transform pair is given by:
and
26
T. Peynot
2D FT: Example
Chapter 4
Example: Obtaining the 2-D Fourier transform of a simple function
27
T. Peynot
Chapter 4
2D Sampling
4.5.3 Two-Dimensional Sampling and the 2-D Sampling Theorem
Sampling in 2-D can be modeled using the sampling function (2-D impulse train):
Multiplying f(t,z) by yields the sampled function
28
T. Peynot
2D Sampling Theorem Chapter 4

f(t,z) is said to be band-limited if its Fourier transform is 0 outside of a rectangle

[- µmax, µmax] and [- νmax, νmax]
2-D sampling theorem:
A continuous, band-limited function f(t,z) can be recovered with no error from a
set of its sample if the sampling intervals are
and
29
T. Peynot
Chapter 4
2D Discrete Fourier Transform （DFT） Filtering in the Frequency Domain
4.5.5 The 2-D Discrete Fourier Transform and its Inverse
2-D Discrete Fourier Transform (DFT) of a digital image f(x,y) of size MxN:
Inverse Discrete Fourier Transform (IDFT):

2D DFT example
2D DFT example
translation
insensitive,
but rotation
sensitive
4.6 Some Properties of the 2-D Discrete Fourier Transform

Relations between Spatial & Frequency Domains
4.6.1 Relationships between Spatial and Frequency Domains
If ΔT and ΔZ are the separations between samples, the separations between the
corresponding discrete, frequency domain variables are:
and
31
Filtering
in Frequency Domain
Chapter 4
2D Convolution
4.6.6 The 2-D Convolution Theorem
2-D circular convolution :
2-D convolution theorem :
and conversely:

Wraparound Error & Padding

T. Peynot
Chapter 4
Data from adjacent periods produce

wraparound error
=> Need to append zeros to both functions
In 2-D, padd the two images array of size

MxN by zeros
Padded images of size PxQ, with:
P ≥ 2M-1
and
Q ≥ 2N-1
More 2D
© 1992–2008 R. C.wraparound/padding
Gonzalez & R. E. Woods
example see Page 281 Figure 4.32

T. Peynot
Chapter 4
Frequency Domain Filtering: Fundamentals

4.7.2 Frequency Domain Filtering Fundamentals
Given a digital image f(x,y) of size MxN, the basic filtering equation is:
NB: product = array multiplication
Filtered (output) image Filter function DFT if the input image

(or filter transfer
function)
NB: For simplification, use functions H(u,v)

that are centered symmetric about their
centre
+ centre F(u,v) multiplying f(x,y) by (-1)x+y

Shifting 2D Frequency Spectrum
M
X1
j2⇡ux/M
f (x) , F (u) = f (x)e
x=0
M
X1
f (x)ej2⇡(u0 x/M ) , F (u u0 ) = f (x)e j2⇡(u u0 )x/M
x=0
M
X1
x j2⇡(u M/2)x/M
f (x)( 1) , F (u M/2) = f (x)e
x=0
Extension to 2D:
f (x, y)( 1)x+y , F (u M/2, v N/2)

T. Peynot
Filtering in Frequency Domain

Chapter 4
Lowpass filter highpass filter
008 R. C. Gonzalez & R. E. Woods

Chapter 4
Steps for Frequency Domain Filtering
4.7.3 Steps for Filtering in the Frequency Domain
1. Given an input image f(x,y) of size MxN, obtain 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 centre 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 centre at
coordinates (P/2,Q/2). Form the product G(u,v)=H(u,v)F(u,v) using array
multiplication.
6. Obtain the processed image:
7. Obtain the final processed result, g(x,y), by extracting the MxN region from
the top, left quadrant of gp(x,y)
T. Peynot
Chapter 4
Steps for Frequency Domain Filtering

4.7.3 Steps for Filtering in the Frequency Domain

2) Padding 3) (-1)x+y
1)
4) DFT
5) Filter H
6) Filtered
image (IDFT) 7) Final
Result
4.8 Image Smoothing using Frequency Domain Filters
Ideal Lowpass Filter

NB: all filter functions are assumed to be discrete functions of size PxQ
4.8.1 Ideal Lowpass Filters T. Peynot
Ideal Lowpass Filter (ILPF) = 2-D lowpass filter that passes without attenuation all
Chapter 4
frequencies within a circle of radius D0 from the origin and “cuts off” all frequencies
outside this circle
4.8.1 Ideal Lowpass Filters

Where:
The point of transition between H(u,v) = 1 and H(u,v) = 0 is called the cutoff frequency
And D(u,v) is the distance between a point (u,v) and the centre of the frequency rectangle:
NB: an ILPF cannot be realized with real electronic components

T. Peynot
Ideal Lowpass Filter (Effect) Chapter 4

r = 10
4.8.1 Ideal Lowpass Filters
Example:
r = 30
r = 60
r = 160 r = 460

Ideal Lowpass: Blurring and Ringing

ng the blurring and “ringing” properties of ILPFs
T. Peynot
Chapter 4
Butterworth Lowpass Filter
4.8.2 Butterworth Lowpass Filters
Transfer function of a Butterworth lowpass filter (BLPF) of order n and with

cutoff frequency at a distance D0 from the origin:

Chapter 4
Butterworth Lowpass
Filtering in Filter
the Frequency Domain
4.8.2 (Spatial)
Butterworth Lowpass Filters
Butterworth Lowpass Filter (Effect)

T. Peynot
Chapter 4
4.8.2 Butterworth Lowpass Filters

T. Peynot
Gaussian Lowpass Filter

Chapter 4
4.8.3 Gaussian Lowpass Filters
Gaussian Lowpass Filters (GLPFs) in two-dimensions:
(σ = measure of spread about the centre)
(D0 = cutoff frequency)

Gaussian Lowpass Filter (Effect)

T. Peynot
Chapter 4
4.8.3 Gaussian Lowpass Filters

T. Peynot
Chapter 4
Highpass Filter
4.9 Image Sharpening using Frequency Domain Filters
Highpass filtering: attenuation of the high-frequency components of the Fourier

transform of the image
As before :
•  Radially symmetric filters
•  All filter functions are assumed to be discrete functions of size PxQ
T. Peynot
A highpass HHP filter can be obtainedChapter

from a given
4 lowpass HLP filter by:
4.9.1 Ideal Highpass Filters
A 2-D Ideal Highpass Filter (IHPF) is defined as:

T. Peynot
Highpass Filters (in Frequency Domain) Chapter 4

Ideal Highpass Filter:
Butterworth Highpass Filter:
Gaussian Highpass Filter:

T. Peynot
Highpass Filters (in Spatial Domain) Chapter 4

Ideal Butterworth Gaussian

Highpass Filter Highpass Filter Highpass Filter

T. Peynot
Highpass Filters (Effect)

Chapter 4
4.9.1 Ideal Highpass Filters
Do = 30 Do = 60 Do = 160
T. Peynot
Chapter 4
Ideal HPF Filtering in the Frequency Domain
4.9.2 Butterworth Highpass Filters
Do = 30 Do = 60Processing
Digital Image Do = 160
T. Peynot
Chapter 4
Butterfly HPF
4.9.3 Gaussian Highpass Filters
Do = 30 Do = 60 Do = 160
Gaussian HPF

9.5 Unsharp Masking, Highboost Filtering and High-Frequency-Emphasis Filtering
High Frequency Emphasis Filter

Expressed in terms of frequency domain computations:
Lowpass filter: T. Peynot
Highpass filter: Chapter 4

High-frequency-emphasis filter
4.9.5 Unsharp Masking, Highboost Filtering and High-Frequency-Emphasis Filtering
More general formulation:

Gaussian highpass
filtering
Where: gives controls of the offset from the origin

controls the contribution of high frequencies
High-frequency Histogram
emphasis filtering Equalisation
R. C. Gonzalez & R. E. Woods

Week4 - Frequency Domain Operation

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

Week4 - Frequency Domain Operation

Uploaded by

Copyright:

Available Formats

EIE4512 - Digital Image Processing

Frequency Domain Operations

• Polar Coordinates '

Unit impulse of a continuous variable t located at t=0 :

(if f(t) continuous at t=0 )

© 1992–2008 R. C. Gonzalez & R. E. Woods

Discrete Unit Impulse

Unit discrete impulse located at x=0 :

© 1992–2008 R. C. Gonzalez & R. E. Woods

Impulse Train Filtering in the Frequency Domain

Impulse train : sum of infinitely many periodic impulses ΔT units apart:

© 1992–2008 R. C. Gonzalez & R. E. Woods

Fourier Transform (FT) of Filtering in the Frequency Domain

Fourier Transform of a continuous function f(t) of a continuous variable t :

Inverse Fourier Transform:

© 1992–2008 R. C. Gonzalez & R. E. Woods

© 1992–2008 R. C. Gonzalez & R. E. Woods

FT of Unit Impulse Chapter 4

Example 2 : Fourier Transform of a unit impulse located at the origin:

Fourier Transform of a unit impulse located at t = t0 :

FT of Impulse Train Chapter 4

Example 2 : Fourier Transform S(µ) of an impulse train with period ΔT :

Also an impulse train, with period 1/ΔT

© 1992–2008 R. C. Gonzalez & R. E. Woods 11

Continuous-Time Convolution Filtering in the Frequency Domain

Fourier Transform pairs:

Spatial Domain Conv <=> Frequency Domain Multiplication

Spatial Domain Multiplication <=> Frequency Domain Conv

© 1992–2008 R. C. Gonzalez & R. E. Woods

1992–2008 R. C. Gonzalez & R. E. Woods

FT of Sampled Function Chapter 4

Superposition of infinitely many

© 1992–2008 R. C. Gonzalez & R. E. Woods

Sampling Theorem Chapter 4

A function f(t) whose Fourier transform is zero

NB: Sampling at:

Recovery with Low-pass Filter Chapter 4

=> Recover f(t) using the inverse FT:

Example: ideal lowpass filter

© 1992–2008 R. C. Gonzalez & R. E. Woods

Effect of under-sampling a function

NB: the effect of aliasing can be reduced

© 1992–2008 R. C. Gonzalez & R. E. Woods

Illustration: sampling a sine wave of period 2s

Consider a finite sequence of samples

Inverse Discrete Fourier Transform (IDFT)

• The duration of the record is

• The entire frequency range is

• The spacing in the frequency domain

Discrete equivalent of the convolution:

© 1992–2008 R. C. Gonzalez & R. E. Woods

Extension to Function of Two Variables

4.5.1 The 2-D impulse and its Sifting Property

Impulse of two continuous variables t and z :

Digital Image Processing

4.5 Extension to Function of Two Variables

4.5.1 The 2-D impulse and its Sifting Property

2-D discrete impulse:

© 1992–2008 R. C. Gonzalez & R. E. Woods 25

4.5.2 The 2-D Continuous Fourier Transform Pair

© 1992–2008 R. C. Gonzalez & R. E. Woods

Example: Obtaining the 2-D Fourier transform of a simple function

© 1992–2008 R. C. Gonzalez & R. E. Woods