EENG 860 Special Topics: Digital Image Processing

Lecture 7: Image Filtering in Frequency

Domain

Dr. Ahmadreza Baghaie

Department of Electrical and Computer Engineering
New York Institute of Technology

Spring 2020

Readings: Chapter 4 (sections 4.7-4.9)

1 / 40
How to read: more relaxed than before.
Table of Content

● Basics of Filtering in the Frequency Domain

● Image Smoothing Using Lowpass Frequency Domain Filters
● Image Sharpening Using Highpass Frequency Domain Filters
● Selective Filtering
● The Fast Fourier Transform (FFT)

2 / 40
2D Sampling and 2D Sampling Theorem

● Function f(t,z) is said to be band limited, if its 2D FT is 0 outside a rectangle

in the frequency domain, by the limits [-umax, umax] and [-vmax, vmax]:
F (u , v)=0 for |u|≥u max and |v|≥v max
● The 2D sampling theorem states that a continuous, band-limited function
f(t,z) can be recovered with no error from a set of its samples if the sampling
rates are:

1
>2u max
ΔT
1
>2 v max
ΔZ

3 / 40
2D Aliasing – Example

● Aliasing occurs in image re-sampling as well.

● Image smoothing before re-sampling can be used to reduce the aliasing.

4 / 40
Summary of DFT Definitions

5 / 40
Summary of DFT Pairs

6 / 40
Fourier Spectrum and Phase Angle – Example

● Spectrum or phase angles; which one is more important?

7 / 40
Basics of Filtering in the Frequency Domain

● The Fourier transform is computed by combination of the values of the signal

in the spatial domain, modified by the values of the exponential terms.
● In general, it is impossible to make direct correspondence between specific
components of an image and its Fourier transform.
● Since frequency is directly related to spatial rates of change, we can
associate frequencies in the Fourier transform with patterns of intensity
variations in an image.
● Filtering in the Fourier domain is based on modification of the Fourier
transform to achieve a specific objective, followed by inverse Fourier
transform to obtain the spatial domain representation of the processed result.

g (x , y)=Real {IDFT [ H (u , v) F (u , v)]}

8 / 40
Steps for Filtering in the Frequency Domain

1) Input image, f(x,y) of size MxN; zero-, mirror- or replicate-pad it to form

fp(x,y) with size PxQ, for P=2M and Q=2N;

2) Multiply fp(x,y) by (-1)(x+y) to center the Fourier transform

3) Compute the DFT, F(u,v), of the image from last step;

4) Construct a real, symmetric filter transfer function, H(u,v), of size PxQ with
center at (P/2, Q/2);
5) Form the product G(u,v)=H(u,v)F(u,v) using element-wise multiplication;

6) Obtain the filtered image of size PxQ by computing IDFT of G(u,v):

(x+ y)
g p ( x , y)=(Real {IDFT [G(u , v)]})(−1)
7) Obtain the final filtered result g(x,y) of the same size as the input image, by
extracting the MxN regions from the top, left quadrant of gp(x,y).

9 / 40
Steps for Filtering in the Frequency Domain

10 / 40
Filtering in the Frequency Domain - Examples

11 / 40
Filtering in the Frequency Domain - Examples

● Two ways to design filters:

– Specify the filter in the spatial domain
– Specify the filter in the frequency domain
● Two ways for implementation:
– Use spatial convolution
– Use DFT and IDFT

12 / 40
Filtering in the Frequency Domain - Examples

13 / 40
Image Smoothing Using Lowpass Frequency
Domain Filters
● Edges and sharp intensity transitions such as noise, contribute to high
frequency components in the frequency domain.
● Smoothing (blurring) is achieved by high-frequency attenuation, or lowpass
filtering.
● Three types of filters will be discussed here:
– Ideal lowpass filter (ILPF)
– Gaussian lowpass filter (GLPF)
– Butterworth lowpass filter (BLPF)
● Butterworth filters are controlled by a parameter called the filter order.
● Large values of the filter order, results in shapes similar to the ideal lowpass
filter, while smaller values result in shapes similar to a Gaussian lowpass
filter.

14 / 40
Ideal Lowpass Filter (ILPF)

● A 2D filter which passes without attenuation all of the frequencies within a

circle of specified radius from the origin, and cuts off all of the frequencies
outside this circle.
● Mathematically, it is defined as:
1 if D(u , v) D0
H (u , v) 
0 if D (u , v)  D0
with D0 as the cutoff frequency and:
1
 2
D(u , v)  (u  P / 2)  (v  Q / 2) 
2 2

15 / 40
Ideal Lowpass Filter (ILPF)

● How to select the cutoff frequency? By specifying the percentage of the

image power that you want the filter to pass.
P−1 Q−1
● Total image power is calculated by: P (u , v )=|F (u , v )|2 ⇒ PT = ∑ ∑ P (u , v )
(u=0) v=0

P (u , v )
● The percentage is calculated as: α =100 ∑ PT
(u , v)∈D0

Fourier spectrum
Rings of radii 10, 30, 60,
160, 460 pixels,
corresponding to 87%,
93%, 96%, 98%, 99% of
the total image power.

16 / 40
Ideal Lowpass Filter (ILPF) - Examples

17 / 40
Ideal Lowpass Filter (ILPF) - Limitations

● The main issue with ILPF is the ringing artifacts.

● To see the spatial representation, we can calculate the IDFT of the ILPF.
● The profile of the IDFT of the ILPF is similar to a sinc function; although it is
actually a Bessel function.
● The inverse proportionality between the width of the filter function in the
frequency domain, and the spread of the width of the lobes in the spatial
function still holds.

18 / 40
Gaussian Lowpass Filter (GLPF)

● The transfer function of a Gaussian lowpass filter (GLPF), with cutoff

frequency at a distance D0 from the center of the frequency rectangle:
D 2 ( u ,v )

2 D02
H (u , v) e
1
 2
with D (u , v)  (u  P / 2)  (v  Q / 2) 
2 2

19 / 40
Gaussian Lowpass Filter (GLPF) - Examples

20 / 40
Butterworth Lowpass Filter (BLPF)

● The transfer function of a Butterworth lowpass filter (BLPF) of order n, with

cutoff frequency at a distance D0 from the center of the frequency rectangle:
1
H (u , v)  2n
 D(u , v) 
1  
 D 0 
1
 2
with D (u , v)  (u  P / 2)  (v  Q / 2) 
2 2

21 / 40
Butterworth Lowpass Filter (BLPF) - Examples

22 / 40
Butterworth Lowpass Filter (BLPF)

● Using higher values of the filter order n, BLPF’s characteristics approach to

that of ILPF.
● Using lower values of the filter order n, BLPF’s characteristics approach to
that of GLPF.
● What about ringing?

23 / 40
Lowpass Filtering – Additional Examples

● Text image processing: smoothing/blurring can be used to bridge small gaps

and repair the broken characters.

24 / 40
Lowpass Filtering – Additional Examples

● Printing and publishing industry: cosmetic processing for smoothing and

creating a softer-look, to reduce sharpness of fine skin lines and small
blemishes

25 / 40
Lowpass Filtering – Additional Examples

● To reduce machine-induced artifacts, such as scan lines in the images; for

example a very high resolution radiometer (VHRR) image.
● Also, it can smooth out small features less than certain sizes.

26 / 40
Image Sharpening Using Highpass Frequency
Domain Filters
● Edges and sharp intensity transitions contribute to high frequency
components in the frequency domain.
● Image sharpening can be achieved by highpass filtering, which attenuates
low-frequency components without disturbing high-frequency components.
● Here, we discuss:
– Ideal highpass filter (IHPF)
– Gaussian highpass filter (GHPF)
– Butterworth highpass filter (BHPF)
– Laplacian in Frequency Domain
– Unsharp Masking and High-Boost Filtering

27 / 40
Ideal, Gaussian and Butterworth Highpass Filters

● The transfer function of these highpass

filters can be computed by subtracting a
lowpass filter transfer function from 1:
H HP (u , v) 1  H LP (u , v)

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

D 2 ( u ,v )

2 D02
● GHPF: H (u, v) 1  e
1
● BHPF: H (u , v)  2n
 D0 
1  
 D(u , v) 
28 / 40
Ideal, Gaussian and Butterworth Highpass Filters

● The spatial kernel can be derived by applying IDFT to the HHP:

h HP ( x , y)=IDFT [ H HP (u , v)]=IDFT [1−H LP (u , v)]= δ (x , y)−h LP ( x , y )

29 / 40
Ideal, Gaussian and Butterworth Highpass Filters -
Examples

30 / 40
Ideal, Gaussian and Butterworth Highpass Filters -
Examples

31 / 40
Laplacian in Frequency Domain

● Remember the Laplacian operator in the spatial domain:

2 2 f 2 f
 f  2  2
x y
● Its Fourier transform can be written as:
H (u , v)  4 2 (u 2  v 2 )
● Given that we center our FTs, we can rewrite it as:
H (u, v)  4 2 [(u  P / 2) 2  (v  Q / 2) 2 ]  4 2 D 2 (u, v)
with
1
 2
D(u , v)  (u  P / 2)  (v  Q / 2) 
2 2

● From the convolution theorem, we have:

2 f  1{H (u , v) F (u , v)}

32 / 40
Image Sharpening Using Laplacian Operator

● Remember image sharpening in the spatial domain:

g ( x, y )  f ( x, y )  c2 f ( x, y )
● In the frequency domain, with c=-1:
g ( x , y ) = IDFT {F (u , v )+cH (u , v ) F (u , v )}
= IDFT {[1+cH (u , v )] F (u , v )}
2 2
= IDFT {[1+4  D (u , v )] F (u , v )}
● Image f(x,y) should be normalized to range [0,1], and the Laplacian needs to
be divided by its maximum to bring its range to approximate [-1,1].

33 / 40
Unsharp Masking and High-Boost Filtering

● Remember the equations for unsharp masking (k=1) and high-boost filtering
(k>1):
g ( x, y )  f ( x, y )  k * g mask ( x, y )
g mask ( x, y )  f ( x, y )  f LP ( x, y )
),(yxfg

● We have:
f LP ( x , y)=IDFT {H LP (u , v ) F (u , v )}⇒ g mask ( x , y )=IDFT {F (u , v)−H LP (u , v ) F (u , v )}
● Therefore:
g ( x , y ) = IDFT {(1+k [1−H LP (u , v)]) F (u , v)}
= IDFT {[1+kH HP (u , v)] F (u , v)}

34 / 40
Selective Filtering

Filter Type Spatial Kernel Frequency Kernel

Lowpass lp( x , y)
),(yxfg
LP (u , v)
Highpass hp(x , y)=δ (x , y)−lp( x , y) HP(u , v)=1− LP(u , v )
Bandreject br (x , y)=lp1 ( x , y)+hp2 ( x , y) BR(u , v)= LP1 (u , v)+ HP2 (u , v)
Bandpass bp( x , y)=δ ( x , y)−br ( x , y) BP(u , v)=1−BR(u , v)

35 / 40
Selective Filtering: Notch Filters

● Notch filters reject (pass) frequencies in a

predefined neighborhood of the frequency
rectangle.
● Notch reject filters are constructed by
product of several highpass filter transfer
functions, with centers translated to the
centers of the notches:
Q
H NR (u , v)= ∏ H k (u , v ) H −k (u , v)
k=1

36 / 40
Fast Fourier Transform (FFT)

● We mentioned before that the 2D DFT is separable into 1D transforms:

M −1 N −1
DFT [f ( x , y )] = ∑ ∑ f ( x , y )e− j 2  (ux / M + vy / N )
x=0 y=0
M −1 N −1
− j 2  ux / M
= ∑e ∑ f ( x , y ) e− j 2  vy / N
x=0 y=0
M −1
= ∑ F ( x , v ) e− j 2  ux / M
x=0
= F (u , v )
● We also mentioned that inverse DFT can be calculated using a forward DFT
algorithm:
M −1 N −1
* * − j 2  (ux / M + vy / N )
MNf ( x , y)= ∑ ∑ F (u , v)e
u=0 v =0

37 / 40
Fast Fourier Transform

● Brute-force implementation of the DFT requires on the order of (MN)2

multiplications and additions; exponentials are only calculated once and
stored as a look-up table.
● For an image of 2048x2048 pixels, this means 17 trillion calculations.
● Fast Fourier Transform (FFT) brings this down to MN log2 MN
multiplications and additions.
● For an image of similar size as before, FFT requires 92 million calculations!
● For an image of size MxM with M=2p, the computational advantage of FFT
over a brute-force implementation of the 1D DFT is:

2p
C ( p)=
p

38 / 40
What is Next?

● Image Degradation/Restoration Process

● Noise Models
● Noise Reduction
● Degradation Function Estimation

39 / 40
Questions?
[email protected]

40 / 40