Image Enhancement

Frequency Domain
Filtering
*Frequency Domain and
Spatial Domain Filtering

*Filtering can be done in Spatial domain or

in Frequency domain

*Filteringin spatial domain is easier to

understand

*However, Filtering in frequency domain is

much faster – especially for large images
*Filtering in Spatial
Domain

Image
Convolutio
n
Filter Mask
Ex.
Averaging,
Laplacian
*Filtering in Frequency
Domain

DFT DFT DFT

Convolution in spatial domain

becomes multiplying in frequency
domain
 IDFT After applying Filter
H(u,v) to the image
Tranform F(u,v) we
g (x,y) get the response
G(u,v)
Filtered
Image To recover the
original image we
 * Correspondence between
Frequency domain and Spatial
domain transform
*Spatial Domain : Filtering is given by
CONVOLUTION

*Frequency Domain : Convolution in FD is

equivalent to MULTIPLICATION

*Filtering can be done in either domain

 Applications of Frequency Selective
Filters

* Lowpass filters (achieve

smoothing/blurring)
* Low frequency in transform relates to slowly
varying intensity components in an image such
as walls of a room or a cloudeless sky
* Highpass filters (enhance sharp
details, removes dc term)
* High frequency are due to sharp transition in
intensity such as edges and noise
* Low frequency and High Frequency
components in an image
 * Mask H(u,v) used for filtering

Transform of
an image

Low pass filter mask High pass filter mask

 *2D mask for Filter
H(u,v)

Low pass filter mask Band pass filter mask High pass filter mask
 *Example of LPF
let’s apply this filter to an actual image and let’s see what we got
Sample image

Low pass filtering

smoothes / blurs the
image
Image in frequency domain

IDFT
X =

F(u,v) x H(u,v) =
G(u,v)
 *LPF and HPF
Low Pass Filter

• The edge content is reduced (blurs the image)

• As the size of the mask grow, more smoothing effect will
take place

High Pass Filter

• The edge content (sharpness) is increased

• As the size of the mask grows , more edge content is
increased
Low Pass Filtering

High Pass Filtering

*Ideal Low Pass
Filter
 Ideal Lowpass filter Mask
• Transfer Function , H(u,v) of a ideal LPF

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

D0
Image
representati u
on of filter
D>D0

v
 Ideal Lowpass filter
• D(u,v) is a distance from the center to a point at (u,v)
• D(u,v)=(u2+v2)1/2
• D0 is cut off frequency in terms of distance from center
• Assuming coordinates of center as (0,0)

D0

D>D0
*Example Image Filters
1. Image matrix
f=

2. Multiply by to center DFT

f’=

f’=

f’=
 f’=

4. Pad the image if DFT is of order 4

New image matrix
fnew=

F=
5. DFT of image matrix
F=T. fnew .T’
*Example Image Filters
F=

H=
6. Choose a Low Pass filter for image

[ ]
000 0
F ′ ′ = F × 𝐻 = 0 1+3 𝑖 3+ 3𝑖 0
0 4 +2 𝑖 6 0
000 0
*Example Image Filters
7. Filtered Image

[ ]
0.87 0.87 0.25 0.25
IDFT ( F ¿¿ ′ ′ )= 𝑓 ′ ′ = − 1 0.62 − 0.12 0.50 ¿
0.37 − 0.12 0 − 0.25
− 0 .25 0 .37 − 0 .12 0
 *For HPF
F=

We retain the outer coefficients

F=

And find the inverse DFT to recover the filtered image

 *Effect of cutoff
frequency (Do)
Do controls the amount of bluring of the image

D0
Power spectrum of DFT of image/filter

* F(u,v)  DFT/FFT of image

* H(u,v)  DFT/FFT of filter

* Power spectrum for each value of u,v

P (u , v) Re{H (u , v)}2  Im{H (u , v)}2

 image Preprocess image to center DFT
2 4 1 6  2  4 1  6
3 2 1 4  3 2  1 4 
f ( x , y )  f ' ( x, y ) ( 1) ( x y ) f ( x , y )  
6 5 3 2  6  5 3  2
   
1 2 4 3  1  2 4  3 
 3 54 j 9 5 4 j 
 9 2 j 1  3  3  10 j 
DFT of f’(x,y) F ' (u , v) 
 7 3 2 j 49 32 j 
 
 9  2 j  3  10 j  3 1 

 9 41 81 41 
 85 1 9 109 
P (u , v) Re{H (u , v)}2  Im{H (u , v)}2 P (u , v)  
 49 13 2401 13 
 
 85 109 9 1 

Power spectrum of an
  9 41 81 41 
 85 1 9 109 
P (u , v)  
 49 13 2401 13 
 
 85 109 9 1 

• Total power= 9+40.96+81+…+ 9+1 = 3056

• Power contained in 2x2 within the center

= 1+9+12.96+2401 =2423.96

• Percentage of total power within 2x2 in the center

=100 (2423.96)/3056
=79%

Power contents of an
 *What is the
Inference of the
power calculations
that we did
 Percentage of total power contained
within D0 from center
D (u 2  v 2 )1 / 2 u and v are row and column numbers of
DFT D0=230
, 99.5%

D0=80,
98.0%

D0=30,
96.4%

D0=15
,
94.6%

Image representation D0=5,

Image of magnitude of DFT 92.0%
of Image
27
Effect Original
500x500 Do = 10

of
lowpas
s filters
with D = 30
o

various Do = 60

values
of
cutoff D = o

freq, 160
Do =
Do 460

28
 Original
500x500 Do = 10

Effect
of Do = 30
lowpas Do = 60

s filters
with
various Do =
values 160
Do =
of Do 460

29
 Original
500x500 Do = 10

Do = 30
Do = 60

Do =
160
Do =
c 460

30
Do controls the amount of bluring of the image
Applications of low pass filtering
• LPF fills gaps of broken lines.

• Uses unsharp masking which uses

blurring as first step

• Cosmetic processing for smoother

and soft looking instead of sharp
original image

• For human faces, it reduces sharp fine

lines and blemishes
LPF reduces sharp fine lines and blemishe

*Lowpass Filtering
Examples (cont…)
 LPF fills gaps of broken lines
444 x 508, Low resolution After Guassian low pass filter
with cut off frequency = 80

*Lowpass Filtering
Examples
 LPF fills gaps of broken lines
444 x 508, Low resolution After Guassian low pass filter
with cut off frequency = 80

*Lowpass Filtering
Examples
 Highpass Filters

H hp (u , v) 1  H lp (u , v)

0, if D(u , v) D0

*Ideal filters : H (u, v) 
1, if D(u, v)  D0
Band pass filters

HBP(u,v)=1- HBR(u,v)
 Band reject filter

Band reject filter Original Image Filtered Image

 band pass filter

Band pass filter Original Image Filtered Image

 *Steps for
Frequency
Domain Filtering

* Pad the MxN image to PxQ and Multiply the input image by (-1)x+y 
PREPROCESSING
* Compute F(u,v)  DFT of f(x,y)
* Multiply F(u,v) with Filter transfer function H(u,v)
* Compute the Inverse DFT
* Obtain the real part, crop the image to MxN and multiply the real part by (-
1)x+y  POSTPROCESSING
 Padded image Multiply by (-1)x+y

image

256×256 Produc
t of
Spectru H and
m of F
image

Centered LPF
After IDFT
and
product Cropped image

Steps for frequency domain

 Padded image Multiply by (-1)x+y

image

256×256 511×511 Produc

t of
Spectru H and
m of F
image

Centered LPF
After IDFT
and
product Cropped image

Steps for frequency domain

 Padded image Multiply by (-1)x+y

image

 

256×256 511×511 Produc

511×511
t of
Spectru H and
m of F
image

Centered LPF
After IDFT
and
product Cropped image

Steps for frequency domain

 Padded image Multiply by (-1)x+y

image

 

256×256 511×511 Produc

511×511
t of
DFTof H and
image F


511×511
Centered LPF
After IDFT
and
product Cropped image

Steps for frequency domain

 Padded image Multiply by (-1)x+y

image

  

256×256 511×511 511×511

H(u,v)x
DFT of F(u,v)
image=  
F(u,v) 

511×511 511×511LPF H(u,v) 511×511

After IDFT
and
product Cropped image

Steps for frequency domain

 Padded image Multiply by (-1)x+y

image

 

256×256 511×511 511×511

H(u,v)x
Spectru
F(u,v),
m of  
image,  
F(u,v)
511×511 511×511 511×511LPF H(u,v) 511×511

IDFT{T(u,v)
} 
511×511 Cropped image

511×511

Steps for frequency domain

 Padded image Multiply by (-1)x+y

image

 

256×256 511×511 511×511

H(u,v)x
Spectru
F(u,v),
m of  
image,  
F(u,v)
511×511 511×511 511×511LPF H(u,v) 511×511

IDFT{T(u,v)
} Post processing,
 Cropped and
511×511
filtered image
511×511 256×256

Steps for frequency domain

*Homomorphic
Filtering
 *Homomorphic
Filtering
• Homomorphic Filtering is used to address the problem of improving
the quality of image that has been acquired under poor
illumination conditions

• Homomorphic filtering uses illumination – reflection model

• An image is formed due to the fact that there is some amount of light
falling on the object and some amount of light that is reflected from the
object
 *illumination – reflection
model
(IR model)
• Illumination I (x,y) : Light falling on the object
• Reflectance R(x,y) : light reflected by the object

• These two combine to form an image

• An image can be modeled as the product of an illumination function

and reflectance function at every point on the image.

𝐼 ( 𝑥 , 𝑦 )=𝐿 ( 𝑥 , 𝑦 ) × 𝑅 (𝑥 , 𝑦 )

Illumination Reflectance
 *Low freq. and High Freq
in a IR model of an
𝑓 ( 𝑥 , 𝑦 ) =𝐼 ( 𝑥 , 𝑦 ) × 𝑅 ( 𝑥 , 𝑦 )
image
Illumination Reflectance

• Illumination is the primary contributor to the dynamic

range and varies slowly in space

• While reflector component represents the details of object

edges and varies rapidly in space

• These characteristics lead to associating the

• LOW FREQUENCIES of the Fourier Transform of an image
with illumination and HIGH FREQUENCIES with
reflectance
*Homomorphic Filtering
• The ideal of Homomorphic Filtering is to separate the
illumination (low frequency) and the reflectance
(high frequency) components
*Homomorphic Filtering
• Now to separate the LOW and HIGH Frequency
we have to convert the image f(x,y) in frequency
domain by applying Fourier Transform

𝑖 . 𝑒 . 𝐹 [ 𝑓 ( 𝑥 , 𝑦 ) ]=𝐹 [ 𝑖 ( 𝑥 , 𝑦 ) × 𝑟 ( 𝑥 , 𝑦 ) ]

• The problem with fourier transform is that the

product of two functions is not seperable

]]
*Procedure for applying
Homomorphic Filtering
STEP 1 :
Apply log function to the image , f(x,y)

STEP 2:
Apply fourier Transform to the Log of these components

𝑍 (𝑢 , 𝑣)= 𝐹 𝑖 (𝑢 , 𝑣)+ 𝐹 𝑟 (𝑢 , 𝑣)
 Step 3: construct a frequency-domain high-pass
filter / Low pass filter [H(u,v)]

𝐹 {𝑙𝑛 [ 𝑓 ( 𝑥 , 𝑦 ) ] }=𝐹 {𝑙𝑛 [ 𝑖 ( 𝑥 , 𝑦 ) ] }+ F {ln [ 𝑟 ( 𝑥 , 𝑦 ) ] }

USE LPF
To extract the USE HPF
Low To extract
Frequency High
component Frequency
component
* Block Diagram of
Homomorphic Filtering
Step 4: Take inverse Fourier Transform to the filtered
component

Step 5: To offset the logarithm applied in Step – 1 apply antilog

function to recover the filtered image

𝑠(𝑥 , 𝑦)
𝑔 ( 𝑥 , 𝑦 )= 𝑒 =𝑖𝑜 ( 𝑥 , 𝑦 ) 𝑟 𝑜 ( 𝑥 , 𝑦)

Here, io and ro are the illumination and the reflectance components of

the filtered output image.