1

Unit - 3
Frequency Domain
Image Enhancement Techniques
 2

Image Enhancement

Spatial Frequency
Domain Domain
methods methods

Direct manipulation Modifying Fourier

of pixel in an image transform of image
 3

Background

 The frequency domain refers to the plane of the

two dimensional discrete Fourier transform of an
image

 The purpose of Fourier transform is to represent

a signal as a linear combination of sinusoidal
signal of various frequencies
 4

Background

 Any function that periodically repeats itself can

be expressed as he sum of sines and/or cosines

different coefficient (Fourier Series)

of different frequencies, each multiplied by a

 Even function that are not periodic can also be

multiplied by a weighting function (Fourier

expressed as the integral of sines and/or cosines

Transform)
 5

1D & 2D
FOURIER TRANSFORMATIONS
 6

One-Dimensional Fourier Transform

(Continues domain)

 f(x) is single value continues function

 Its Fourier Transform F(u) can be given as:

①
where

 f(x) can be obtain from F(u) by inverse Fourier

Transform:

②
 7

Two-Dimensional Fourier Transform

(Continues domain)

 These equations make Fourier Transform Pair

 These equations can easily extended to two
variables:

③
 Similarly for inverse transform:

④
 8

One-Dimensional Fourier Transform

(Discrete domain)

 f(x) is single value discrete function

where x = 0 , 1 , 2 , … , M-1
 Its Fourier Transform F(u) can be given as:

⑤
 This is called DFT (Discrete Fourier Transform)
 Inverse DFT can be given as:

⑥
 9

One-Dimensional Fourier Transform

(Discrete domain)

 In previous equations DFT is multiply by while

inverse DFT is not

 If you want to multiply both equations by some

constant then both need to be multiply by

 Altimetry the multiplication of both constant (in

DFT and in inverse DFT) must be
 10

One-Dimensional Fourier Transform

(Discrete domain)

 Euler’s formula

 Substitute this expression into DFT equation

⑦
 11

One-Dimensional Fourier Transform

(Discrete domain)

 Equation ⑤ and ⑦ are complex quantities

 From analysis of complex numbers it is found
that, sometimes it is convenient to express F(u)
in polar domain
( in polar domain )

Real part
Magnitude/spectrum Imaginary Phase angle / Phase spectrum
of Fourier transform part of Fourier transform
 12

One-Dimensional Fourier Transform

(Discrete domain)

Power spectrum / Spectral density

 13

One-Dimensional Fourier Transform

(Discrete domain)

 Form interval [0,M-1] samples can be taken at

equally spaced or arbitrary

 denotes first point of sample

 First value of sampled function is

 Next sample has taken at fixed interval

 Next value of sample function is

 14

One-Dimensional Fourier Transform

(Discrete domain)

 kth sample given by

 Final sample given by

 So is understood as

 So is understood as
 15

Example
F(u)
f(x)
AK
M

K points
A

x u
M points
M points

A discrete function Fourier spectrum of it

of M points
 16

Two-Dimensional Fourier Transform

(Discrete domain)

 These equations make Fourier Transform Pair

 These equations can easily extended to two
variables:

⑩
 Similarly for inverse transform:

⑪
 17

Two-Dimensional Fourier Transform

(Discrete domain)

 Fourier spectrum, phase angle and power

spectrum given as:

⑭
 18

Two-Dimensional Fourier Transform

(Discrete domain)

 It is common practice to multiply the input

image by (-1)x+y prior to computing the Fourier
transform

 Due to properties of exponential

Fourier Transform
of argument
 19

Two-Dimensional Fourier Transform

(Discrete domain)

 So origin of Fourier Transform is located at

u=M/2 and v=N/2

 Means, multiplying by shifts the origin of to

frequency coordinates

 This will be the center point of area occupied by

DFT (M×N size)

 We refer to this area of frequency domain as the

frequency rectangle
 20

Two-Dimensional Fourier Transform

(Discrete domain)

 In computer implementation we need to take

 So the actual center of DFT will be at

 21

Two-Dimensional Fourier Transform

(Discrete domain)

 Place in ⑩

⑯
 is image then Fourier Transform of origin is
equal to the average gray level of the image
 22

Two-Dimensional Fourier Transform

(Discrete domain)

 is real, its Fourier transform is conjugate

symmetric

∴ ⑰

∴ ⑱

 Sampling in spatial and frequency domain has

relationship:

⑲
 23

FILTERING IN
FREQUENCY DOMAIN
 24

Steps of filtering in Frequency Domain

1. Multiply the input image by (-1)x+y to center the

transform

2. Compute

3. Multiply by a filter function

4. Compute the inverse DFT of the result in (3)

5. Obtain the real part of the result in (4)

6. Multiply the result in (5) by (-1)x+y

 25

Steps of filtering in Frequency Domain

Filter Inverse
Fourier
Function Fourier
Transform
H(u,v) Transform

Pre- Post-
Processin Processin
g g

f(x,y) g(x,y)

Input Enhanced
image image
 26

Steps of filtering in Frequency Domain

𝐹 (𝑢,𝑣)=ℑ[𝑓 ( 𝑥 ,𝑦)]
20

Fourier
transform

Input image f(x,y)

Fourier Transform Input
image
 27

Steps of filtering in Frequency Domain

𝐺 ( 𝑢 , 𝑣 )= 𝐻 ( 𝑢 , 𝑣 ) 𝐹 ( 𝑢 , 𝑣 )
21

Filter
function

Fourier transform of
Input image

output image g(x,y)

Fourier transform of
 28

Steps of filtering in Frequency Domain

−1
22

( )
𝑔(𝑥,𝑦)=ℑ [𝐺 𝑢,𝑣 ]
Filtered
Output
image
Inverse Fourier
transform

output image g(x,y)

Fourier transform of
 29

Notch Filter

 Force average value of an image to zero

 Set F(0,0) to zero in frequency

 Do this by multiplying all values of F(u,v) by

filter function:

23
 30

Notch Filter

 This filter
 Set F(u,v) to zero
 Other frequency components remain untouched

 Processed image obtained by inverse Fourier

transform of

 This filter is called notch filter because

it is a constant function with a hole (notch) at
the origin
 31

Notch Filter

SEM image of Result of filtering

a damaged IC With Notch filter
 32

LPF & HPF

 LPF ( Low pass filter )

 Reduces high frequencies

 Passing low frequencies
 Smoothing images

 HPF ( High pass filter )

 Reduces low frequencies

 Passing high frequencies
 Sharpening images
 33

LPF & HPF

2D LPF
Result
function
of
LPF

Result
2D HPF of
function HPF
 34

Correspondence between filtering

Spatial domain & Frequency domain

 Convolution of and of size M×N

24
 35

Correspondence between filtering

Spatial domain & Frequency domain
 Let and denote the Fourier transformation of an
respectively then Fourier transform pair is:

25
 36

Correspondence between filtering

Spatial domain & Frequency domain
 Let and denote the Fourier transformation of an
respectively then Fourier transform pair is:

Fourier transform

Inverse Fourier transform

 37

Correspondence between filtering

Spatial domain & Frequency domain
 Let and denote the Fourier transformation of an
respectively then Fourier transform pair is:

Fourier transform

Inverse Fourier transform

 Visa versa of previous equation is:
 38

SMOOTHING
FREQUENCY DOMAIN FILTERS
 39

Basic

 Smoothing (blurring) is achieved by

attenuating a specified range of high-frequency
components in the transform of a given image

 Done by Low Pass Filters (LPF)

 3 types of LPF need to study

 Ideal
 Butterworth
 Gaussian
 40

1. Ideal Lowpass Filters ( ILPF )

 Simplest low pass filter

 Cuts off all high frequency components of FT
> D0 (distance from origin)

26

Non negative

Distance from point

(u,v) to origin
 41

1. Ideal Lowpass Filters ( ILPF )

 If image size is M×N then

center of the frequency rectangle is at

 Distance of point (u,v) from origin is

27
 42

1. Ideal Lowpass Filters ( ILPF )

Frequency in this Frequency in this

range will pass Range is attenuated

Cutoff
frequency

3D perspective H(u,v) Filter radial

plot of H(u,v) as image cross section
 43

1. Ideal Lowpass Filters ( ILPF )

 Cutoff frequency

 The point of transition between and in ideal low pass

filter cross section
 44

1. Ideal Lowpass Filters ( ILPF )

 Can compare the behavior of ILPF with different

cutoff frequency
 One way to establish a set of standard cutoff
frequency is to compute circles that enclose
specified amounts of total image power PT

28

 P(u,v) is given in Eq. ⑭

 45

1. Ideal Lowpass Filters ( ILPF )

 Transform is a circle with radius r and origin as

center
 Then it encloses ⍺ percentage of power

29

 Summation is taken over the values of (u,v) that

lie inside the circle or on boundary
 46

1. Ideal Lowpass Filters ( ILPF )

 47

1. Ideal Lowpass Filters ( ILPF )

 48

2. Butterworth Lowpass Filters (BLPF)

 Transform function of Butterworth filter is

30

 No sharp discontinuity like ILPF that establishes

a clear cutoff between passed and cutoff
frequency
 49

2. Butterworth Lowpass Filters (BLPF)

 50

2. Butterworth Lowpass Filters (BLPF)

 51

2. Butterworth Lowpass Filters (BLPF)

 52

3. Gaussian Lowpass Filter ( GLPF )

 Transform function of Gaussian filter is

31

 is a measure of the spread of the Gaussian

curve
 Rewrite the above equation by letting =D0
32
 53

3. Gaussian Lowpass Filter ( GLPF )

 54

3. Gaussian Lowpass Filter ( GLPF )

 55

3. Gaussian Lowpass Filter ( GLPF )

 56

3. Gaussian Lowpass Filter ( GLPF )

 57

3. Gaussian Lowpass Filter ( GLPF )

 58

SHARPENING
FREQUENCY DOMAIN FILTERS
 59

Basic

 Sharpening is achieved by
attenuating a specified range of low-frequency
components in the transform of a given image

 Done by High Pass Filters (HPF)

 60

1. Ideal Highpass Filter ( IHPF )

 Ideal Highpass Filter ( IHPF ) is defined as

33
 61

1. Ideal Highpass Filter ( IHPF )

 62

2. Butterworth Highpass Filter (BHPF)

 Transfer function of the BHPF of order n and with

cutoff frequency D0 is given as

34
 63

2. Butterworth Highpass Filter (BHPF)

 64

3. Gaussian Highpass Filter ( GHPF )

 Transfer function of the GHPF with cutoff

frequency D0 is given as

35
 65

3. Gaussian Highpass Filter ( GHPF )

 66

4. Laplacian in frequency domain

 Fourier transform of nth order derivation of

function with one variable is given as :

36
 67

4. Laplacian in frequency domain

 Laplacian work with 2nd order derivation & our

image is having two variables x and y so the
equation will be given as :

37
 68

4. Laplacian in frequency domain

 We second order derivation of function as

above equation can be written as:

38

39

 Center of frequency domain is at ( M/2 , N/2 )

40
 69

4. Laplacian in frequency domain

 From 38 and 40

41

42

43
 70

4. Laplacian in frequency domain

 In spatial domain Generate enhanced image by

subtracting Laplacian image from original image

 In frequency domain task is done by taking one

filter:

44
 71

4. Laplacian in frequency domain

 Enhanced image is given as

45
 72

HOMOMORPHIC FILTERING
 73

Basic idea

 Method based on special case of a class of

system known as Homomorphic System

 Approach is the separation of the illumination

and reflection components

 The Homomorphic Filter Function H(u,v) can

then operate on these components separately
 74

Homomorphic Filtering

 Image f(x,y) can be expressed as two

components (illumination and reflection)
46

 Fourier transform of the product of two function

is not separable

 So need to follow the following steps to get

enhanced image
 75

Homomorphic Filtering

f(x,y) g(x,y)

ln exp

DFT (DFT)-1

H(u,v)
 76

Homomorphic Filtering

1) Find ln (natural log) of f(x,y)

47

2) Find Fourier transform

48
 77

Homomorphic Filtering

3) Process Z(u,v) with filter H(u,v)

49

4) Find inverse Fourier transform

50
 78

Homomorphic Filtering

5) Find exp of s(x,y)

51
 79

END
OF
UNIT - 3