11/11/2015

University of Technology and Education

Faculty of Electrical & Electronic Engineering

Lecture:
IMAGE PROCESSING
Chapter 4:
Image Filtering

1
Nguyen Thanh Hai, PhD

University of Technology and Education

Faculty of Electrical & Electronic Engineering
Image Filtering

INTRODUCTION
-Image Filtering allows to apply various effects on images in order to
filter noise or to smooth image.
-The 2D filter is a 2D filter matrix

* Filtering in the frequency domain

- Lots of noises: Gauss, impulse, etc.
- Filters: lowpass, highpass, bandpass.
* Filtering in the spatial domain
- Linear filtering is a weighted mask used to express the effects
of the filter on each pixel of the image.

2
Nguyen Thanh Hai, PhD

1
 11/11/2015

University of Technology and Education

Faculty of Electrical & Electronic Engineering
Image Filtering
IMAGE FILTERING IN THE SPATIAL DOMAIN

- A spatial filter means that a typical pixel in image is

calculated based on neighborhood pixels. The filter is
moved to all image pixels to create the output image.

Fig 4.1 describes a 3x3 spatial filter. Assume that one needs
to calculate a image pixel
,  in an image, the
corresponding pixel of output image 
,  is calculated by
the sum of products of coefficients (values) between the filter
and image.

3
Nguyen Thanh Hai, PhD

University of Technology and Education

Faculty of Electrical & Electronic Engineering
Image Filtering

Fig. 4.1. Diagram of the 3x3 spatial

filter and on image.

4
Nguyen Thanh Hai, PhD

2
 11/11/2015

University of Technology and Education

Faculty of Electrical & Electronic Engineering
Image Filtering
IMAGE FILTERING IN THE SPATIAL DOMAIN
- Notice: the center coefficient of filter is (0,0), corresponding to the
image pixel considered
,  . With a filter of × , assume that =
2 + 1 and = 2 + 1, in which  and  are positive integers. The
size of the odd and smallest filter is 3 × 3).

- In general, the × spatial filter used on the  ×  input image to

produce the output image (
, ) is expressed as follows:
 

(
, ) =    ,  
+ ,  + 
 

(
, ) =  −1, −1 
− 1,  − 1 +  −1,0 
− 1, 

+. . . +  0,0 
,  +. . . +  1,1 (
+ 1,  + 1)

5
Nguyen Thanh Hai, PhD

University of Technology and Education

Faculty of Electrical & Electronic Engineering
Image Filtering
IMAGE FILTERING IN THE SPATIAL DOMAIN
Correlation and convolution in spatial domain

- Correlation in image processing is to move a window (filter) on image

and calculate the sum of products at every position.
- In convolution, except rotating the filter of 180# .
- Notice: convolution means that a filter moves and rotate a degree of
180# and produce result rotated 180# as described in Figs 4.2(a) and
4.2(b).

- Fig. 4.2. is description of correlation and convolution methods

between the 1D signal and the unit impulse. Notice of transposing.

6
Nguyen Thanh Hai, PhD

3
 11/11/2015

University of Technology and Education

Faculty of Electrical & Electronic Engineering
Image Filtering

Fig. 4.2(a) correlation

method.

Original is the 1st

pixel, to correlate to 5
values of w, one can
add zeros as Step-3
and then shift 1 by 1
from left to right.
Every correlation is
calculated by the sum
of products of every
point and in this case,
12 times shifted as
result of Step-6

7
Nguyen Thanh Hai, PhD

University of Technology and Education

Faculty of Electrical & Electronic Engineering
Image Filtering

Fig. 4.2(b) convolution

method

Original is the 1st pixel, to

convolute to 5 values of w,
one can add zeros as Step-
10 and rotate the window w
with a degree of 180 and
then shift 1 by 1 from left to
right. It is calculated by the
sum of products of every
point and in this case, 12
times shifted and then as
result of Step-14

8
Nguyen Thanh Hai, PhD

4
 11/11/2015

University of Technology and Education

Faculty of Electrical & Electronic Engineering
Image Filtering
In general, the correlation expression is described as below equation, the
filter 
,  with the size of × , the image 
,  and called

, 
(
, )
 


, 
(
, ) =    ,  
+ ,  + 
 

Similarly, the convolution of 
,  and 
,  , called 
,  (
, ),

is expressed as follows:
 


,  (
, ) =    ,  
− ,  − 
 

The symbol minus means that the image  is rotated 180 degrees.

9
Nguyen Thanh Hai, PhD

University of Technology and Education

Faculty of Electrical & Electronic Engineering
Image Filtering

(a) (b)

(b)
(d)

Fig 4.3(a) in similarity to Fig 4.2(a), correlation of an

image f and the 3x3 filter w, result as (e) 10
Nguyen Thanh Hai, PhD

5
 11/11/2015

University of Technology and Education

Faculty of Electrical & Electronic Engineering
Image Filtering

(e) (f)

(g)
(h)

Fig 4.3(b) in similarity to Fig 4.2(b), convolution of

an image f and the 3x3 filter w, result as (h) 11
Nguyen Thanh Hai, PhD

University of Technology and Education

Faculty of Electrical & Electronic Engineering
Image Filtering
Ex: Using the correlation and convolution methods to calculate
the output image at the position g(2,1).
y: 0 to 4 y

1 1 1 2 2
t: -1 to -1
1 1 1 2 2
x: 0 to 4 1 1 1 2 2 1 1 1 x 12

2 2 2 2 2 s: -1 to 1 1 1 1
2 2 2 2 2 1 1 1

Input image Mask (filter) Output image

g(x,y)=g(2,1)
g (2,1) = ω (- 1,-1) f (1,0 ) + ω (- 1,0 ) f (1,1) + ω (- 1,1) f (1,2 )
+ ω (0,-1) f (2,0 ) + ω (0,0 ) f (2,1) + ω (0,1) f (2,2 )
+ ω (1,-1) f (3,0 ) + ω (1,0 ) f (3,1) + ω (1,1) f (3,2 ) = 12 12
Nguyen Thanh Hai, PhD

6
 11/11/2015

University of Technology and Education

Faculty of Electrical & Electronic Engineering
Image Filtering
IMAGE FILTERING IN THE SPATIAL DOMAIN
Some functions in MATLAB

- Ones: matrix 1.
- imfilter(f,h,'Boundary Options','OutputSize','Correlation and Convolution
Options')
- 'Boundary Options' : allow to select a boundary of the output image
- 'Output Size‘: select the output size with 'Correlation and
Convolution Options'.
- More functions in Table. 4.1 of Image Processing book

13
Nguyen Thanh Hai, PhD

University of Technology and Education

Faculty of Electrical & Electronic Engineering
Image Filtering
Some filters expressed by mathematic methods:

* Gaussian (
− n12 + n22 )
hg (n1 , n2 ) = e 2σ 2

hg (n1 , n2 )
or h(n1 , n2 ) =
∑n ∑n h
1 2 g

* Laplacian
∂2 ∂2
∇2 = +
∂x 2 ∂y 2

 α 1−α α 
1 + α 1+α 1+α 
or 1 − α −4 1−α 
 
1 + α 1+α 1+α 
 α 1−α α 
1 + α 1+α 1 + α  14
Nguyen Thanh Hai, PhD

7
 11/11/2015

University of Technology and Education

Faculty of Electrical & Electronic Engineering
Image Filtering

* Log (Laplacian of Gaussian)

h(n1 , n2 ) =
(n
2
1 )
+ n22 − 2σ 2 hg (n1 , n2 )
2πσ 6
∑n ∑n h
1 2 g

* Prewitt
1 1 1
h =  0 0 0 
− 1 − 1 − 1

* Sobel
1 2 1
h =  0 0 0 
− 1 − 2 − 1

15
Nguyen Thanh Hai, PhD

University of Technology and Education

Faculty of Electrical & Electronic Engineering
Image Filtering
Ex 4.1: Express image filtering using filters in toolbox of MATLAB

clear all;
f=imread('cameraman.bmp');
ha=fspecial('average');
ga=imfilter(f,ha,'replicate');
hd=fspecial('disk'); hm=fspecial('motion');
gd=imfilter(f,hd,'replicate'); gm=imfilter(f,hm,'replicate');
hg=fspecial('gaussian'); hp=fspecial('prewitt');
gg=imfilter(f,hg,'replicate'); gp=imfilter(f,hp,'replicate');
hlap=fspecial('laplacian'); hs=fspecial('sobel');
glap=imfilter(f,hlap,'replicate'); gs=imfilter(f,hs,'replicate');
hlog=fspecial('log');
glog=imfilter(f,hlog,'replicate');

16
Nguyen Thanh Hai, PhD

8
 11/11/2015

University of Technology and Education

Faculty of Electrical & Electronic Engineering
Image Filtering

(a) (b) (c)

(d) (e) (f)

17
Nguyen Thanh Hai, PhD

University of Technology and Education

Faculty of Electrical & Electronic Engineering
Image Filtering

(g) (h) (i)

Fig 4.4. Express images with different filters

(a) Original image; (b) Average; (c) Disk; (d) Gaussian; (e)
Laplacian; (f) Log; (g) Motion; (h) Prewitt; (i) Sobel

18
Nguyen Thanh Hai, PhD

9
 11/11/2015

University of Technology and Education

Faculty of Electrical & Electronic Engineering
Image Filtering
IMAGE FILTERING IN THE FREQUENCY DOMAIN
- Some applications in frequency domain, one often uses the FT.
- Image filtering in the frequency domain

19
Nguyen Thanh Hai, PhD

University of Technology and Education

Faculty of Electrical & Electronic Engineering
Image Filtering
IMAGE FILTERING IN THE FREQUENCY DOMAIN

According to diagram of Fig above, frequency filtering can be expressed

in MATLAB using the following functions:

function g=dftfilt(f,H)
F=fft2(f,size(H,1),size(H,2));
G=H.*F;
g=ifft2(G);
f=real(g);
g=g(1:size(f,1),1:size(f,2))
end

20
Nguyen Thanh Hai, PhD

10
 11/11/2015

University of Technology and Education

Faculty of Electrical & Electronic Engineering
Image Filtering
IMAGE FILTERING IN THE FREQUENCY DOMAIN
- H is the impulse response of filter with frequency, assume that its size
is 4 times of the input image. In the spatial domain, one uses
convolution, but in the frequency, the multiplication is used. After
filtering, the output values (matrix) of the filter will be converted using
IFT to obtain the image with the same size of the original image.

- Building the frequency filter using functions in MATLAB as freqz2

with FIR (Finite Impulse Response)
- H=freqz2(h,R,C)
In which h is the 2D spatial filter, H is the 2D frequency filter
response, R and C are the number of rows and columns of H.

21
Nguyen Thanh Hai, PhD

University of Technology and Education

Faculty of Electrical & Electronic Engineering
Image Filtering
IMAGE FILTERING IN THE FREQUENCY DOMAIN

Ex 4.2: Express image filtering in the

spatial domain using Sobel and in the
frequency domain using DFT

clear all;
f=imread('cameraman.tif');
h=fspecial('sobel');
sizeH=size(f)*2;
H=freqz2(h,sizeH);
H1=ifftshift(H);
gd=dftfilt(f,H1);
gs=imfilter(double(f),h);
d=abs(gd-gs);
gd=uint8(gd);imshow(gd);
gd=uint8(gd);imshow(gs);
22
Nguyen Thanh Hai, PhD

11
 11/11/2015

University of Technology and Education

Faculty of Electrical & Electronic Engineering
Image Filtering
IMAGE FILTERING IN THE FREQUENCY DOMAIN
Checking homologous between two methods:
$ =  − $

In which gs is the image filtered in the spatial and gd is filtered in

the frequency, results as in Fig. 4.6

Fig. 4.6. Sobel filter

(a). in the frequency
(b) in the spatial
(c) differential max $ = 3.4106 × 10*+
23
Nguyen Thanh Hai, PhD

University of Technology and Education

Faculty of Electrical & Electronic Engineering
Image Filtering
IMAGE FILTERING IN THE FREQUENCY DOMAIN
Build function dftuv for calculation of distance with grid of .

Ex of an image with 8 × 5 :
function [U,V]=dftuv(m,n)
u=0:(m-1); [U,V]=dftuv(8,5);
v=0:(n-1); D=U.^2+V.^2
idx=find(u>m/2); D=
u(idx)=u(idx)-m; 0 1 4 4 1
idy=find(v>n/2); 1 2 5 5 2
v(idy)=v(idy)-n; 4 5 8 8 5
[V,U]=meshgrid(v,u); 9 10 13 13 10
end 16 17 20 20 17
9 10 13 13 10
4 5 8 8 5
1 2 5 5 2
24
Nguyen Thanh Hai, PhD

12
 11/11/2015

University of Technology and Education

Faculty of Electrical & Electronic Engineering
Image Filtering
IMAGE FILTERING IN THE FREQUENCY DOMAIN

Using function fftshift (D) move coordinate to the center of the

frequency rectangular image (frequency retangle) and the result:

ans =
20 17 16 17 20
13 10 9 10 13
8 5 4 5 8
5 2 1 2 5
4 1 0 1 4
5 2 1 2 5
8 5 4 5 8
10 9 10 13

25
Nguyen Thanh Hai, PhD

University of Technology and Education

Faculty of Electrical & Electronic Engineering
Image Filtering
Lowpass Filter: blurring image
* Ideal LowPass Filter - ILPF) : the transfer functions is described as:
1 1 ., / ≤ 1#
- ., / = 0
0 1 ., / > 1#

where 1# is the cut-off frequency with non-negative value, 1 ., / is the

distance from ., / to the filter center. Orbit of 1 ., / = 1# is a circle.

* Butterworth LowPass Filter - BLPF) with n orders, the cut-off

frequency 1# , the transfer functions is described as:

1
- ., / =
1 + 1 ., / /1# 56

Often define the cut-off frequency at positions of - ., / = 0.5 (reduce

50% compared with the max value 1) when 1 ., / = 1# .

26
Nguyen Thanh Hai, PhD

13
 11/11/2015

University of Technology and Education

Faculty of Electrical & Electronic Engineering
Image Filtering
IMAGE FILTERING IN THE FREQUENCY DOMAIN

Transfer function of the Gaussian Lowpass Filter (GLPF) is expressed

as follows:
9 :,; ⁄5< 9
- ., / = 7 8

In which > is the standard deviation. If > = 1# , we have :

- ., / = 7 8
9 :,; ⁄58?9

When 1 ., / = 1# , the value at the cut-off frequency is 0.0607. Fig.

4.7 describes the 3-D shapes of the frequency response of every filter.

27
Nguyen Thanh Hai, PhD

University of Technology and Education

Faculty of Electrical & Electronic Engineering
Image Filtering

Fig. 4.7. left column: 3-D shape of the window filter; middle
columns: images of the filters; right one: the frequency response 1#
of the ideal filters. First row: the ideal lowpass filter; middle one: the
Butterworth lowpass filter; last one: the Gaussian lowpass filter 28
Nguyen Thanh Hai, PhD

14
 11/11/2015

University of Technology and Education

Faculty of Electrical & Electronic Engineering
Image Filtering
EX: Build a lowpass filter function as
lpfilter with 3 basic filter.

function [H,D]=lpfilter(type,M,N,DO,n) end

[U,V]=dftuv(M,N); H=1./(1+(D./DO).^(2*n));
D=sqrt(U.^2+V.^2); case 'gaussian'
switch type H=exp(-
case 'ideal' (D.^2)./(2*(DO^2)));
H=double(D<=DO); otherwise
case 'btw' error('Unknown filter
if nargin==4 type');
n=1; end
Some functions in toolbox:
- mesh(H): express information in 3D such as x=1:M và y=1:N, with
[M,N]=size(H).
- mesh(H(1:k:end,1:k:end))
- colormap([0 0 0]): black-white
29
Nguyen Thanh Hai, PhD

University of Technology and Education

Faculty of Electrical & Electronic Engineering
Image Filtering
IMAGE FILTERING IN THE FREQUENCY DOMAIN

Ex 4.3: express the plot using mesh axis([0 50 0 50 0 1]);

for the Butterworth lowpass filter with saveas(gcf, 'Hinh3.7b.tif', 'tif');
different orders. BLPF3=fftshift(lpfilter('btw',500,500,
50,3));
clear all; figure;mesh(BLPF3(1:10:500,1:10:5
BLPF1=fftshift(lpfilter('btw',500,500,5 00));
0,1)); axis([0 50 0 50 0 1]);
mesh(BLPF1(1:10:500,1:10:500)); saveas(gcf, 'Hinh3.7c.tif', 'tif');
axis([0 50 0 50 0 1]);
saveas(gcf, 'Hinh3.7a.tif', 'tif'); BLPF4=fftshift(lpfilter('btw',500,500,
50,4));
BLPF2=fftshift(lpfilter('btw',500,500,5 figure;mesh(BLPF4(1:10:500,1:10:5
0,2)); 00));
figure;mesh(BLPF2(1:10:500,1:10:50 axis([0 50 0 50 0 1]);
0)); saveas(gcf, 'Hinh3.7d.tif', 'tif');
30
Nguyen Thanh Hai, PhD

15
 11/11/2015

University of Technology and Education

Faculty of Electrical & Electronic Engineering
Image Filtering
IMAGE FILTERING IN THE FREQUENCY DOMAIN

BLPF3=fftshift(lpfilter('btw',500,500,50,3));
figure;mesh(BLPF3(1:10:500,1:10:500));
axis([0 50 0 50 0 1]);
saveas(gcf, 'Hinh3.7c.tif', 'tif');

BLPF4=fftshift(lpfilter('btw',500,500,50,4));
figure;mesh(BLPF4(1:10:500,1:10:500));
axis([0 50 0 50 0 1]);
saveas(gcf, 'Hinh3.7d.tif', 'tif');

31
Nguyen Thanh Hai, PhD

University of Technology and Education

Faculty of Electrical & Electronic Engineering
Image Filtering
IMAGE FILTERING IN THE FREQUENCY DOMAIN

(a) Order 1 (b) Order 2

(c) Order 3 (d) Order 4

Fig 4.8. Plots of the Butterworth lowpass filters with different orders.

32
Nguyen Thanh Hai, PhD

16
 11/11/2015

University of Technology and Education

Faculty of Electrical & Electronic Engineering
Image Filtering
IMAGE FILTERING IN THE FREQUENCY DOMAIN
Ex 4.4: express the lowpass filter with
different cut-off frequencies and
compare results
D03=3/5*size(f,1);
clear all; [M,N]=size(f);
f=imread('cameraman.bmp'); [H,D]=lpfilter('ideal',2*M,2*N,D03);
g=dftfilt(f,H);
D01=1/5*size(f,1);
[M,N]=size(f); D04=4/5*size(f,1);
[H,D]=lpfilter('ideal',2*M,2*N,D01); [M,N]=size(f);
g=dftfilt(f,H); [H,D]=lpfilter('ideal',2*M,2*N,D04);
g=dftfilt(f,H);
D02=2/5*size(f,1);
[M,N]=size(f); D05=size(f,1);
[H,D]=lpfilter('ideal',2*M,2*N,D02); [M,N]=size(f);
g=dftfilt(f,H); [H,D]=lpfilter('ideal',2*M,2*N,D04);
g=dftfilt(f,H); 33
Nguyen Thanh Hai, PhD

University of Technology and Education

Faculty of Electrical & Electronic Engineering
Image Filtering

C 5C
(a) Original image (b) 1# = (c) 1# =
D D

+C
(d) 1# = (e) 1# =
EC
(f) 1# = 
D D

Fig 4.9. cameraman A × B image after the ideal lowpass filter with
different cut-off frequencies
34
Nguyen Thanh Hai, PhD

17
 11/11/2015

University of Technology and Education

Faculty of Electrical & Electronic Engineering
Image Filtering

Ideal highpass filter

The ideal highpass filter is often implemented to filter image
sharper by eliminating the low frequencies
If the transfer of the lowpass filter is -FG ., / , we have the highpass filter
as follows:
-HG ., / = 1 − -FG ., /

Build the highpass filter function using Matlab

function H=hpfilter(type,M,N,D0,n)
if nargin==4
n=1;
End
Hlp=lpfilter(type,M,N,D0,n);
H=1-Hlp;
end
35
Nguyen Thanh Hai, PhD

University of Technology and Education

Faculty of Electrical & Electronic Engineering
Image Filtering
IMAGE FILTERING IN THE FREQUENCY DOMAIN

Ex 4.5: Express the highpass filter using the

function mesh for image.

clear all;close all;clc

ILPF=fftshift(hpfilter('ideal',500,500,50));
mesh(ILPF(1:10:500,1:10:500));
axis([0 50 0 50 0 1]);

BLPF=fftshift(hpfilter('btw',500,500,50));
mesh(BLPF(1:10:500,1:10:500));
axis([0 50 0 50 0 1]);

GLPF=fftshift(hpfilter('gaussian',500,500,50));
mesh(GLPF(1:10:500,1:10:500));
axis([0 50 0 50 0 1]);
36
Nguyen Thanh Hai, PhD

18
 11/11/2015

University of Technology and Education

Faculty of Electrical & Electronic Engineering
Image Filtering
IMAGE FILTERING IN THE FREQUENCY DOMAIN

(b) (c)
(a)

(d) (e) (f)

Fig. 4.10. Express the highpass filter. Top row: mesh with the 3-D
shape; bottom row: 2D images of the mesh; left column: the ideal
highpass filter, middle columns: the Butterworth highpass filter, right
one: the Gaussian highpass filter
37
Nguyen Thanh Hai, PhD

University of Technology and Education

Faculty of Electrical & Electronic Engineering
Image Filtering
IMAGE FILTERING IN THE FREQUENCY DOMAIN
- The highpass filter is normalized to be zero to the DC components.
Thus, this will reduce the average value of the image onto zero.
- For improvement, the DC offset components are added to the
highpass filter.
- The high frequency emphasis filtering with offset is expressed as
follows:

-HIJ ., / =  + -HG ., /
In which  is the offset,  denotes the multiplication constant and
-HG ., / describes the transfer of the highpass filter.

38
Nguyen Thanh Hai, PhD

19
 11/11/2015

University of Technology and Education

Faculty of Electrical & Electronic Engineering
Image Filtering
IMAGE FILTERING IN THE FREQUENCY DOMAIN

Ex 4.6: express the high frequency emphasis

filtering using Matlab.

clear all;
f=imread('satellite.tiff');
[M N]=size(f);
D0=0.1*size(f,1);
HBW=hpfilter('btw',2*M,2*N,D0);
H=0.5+2*HBW;
gbw=dftfilt(f,HBW);
gbw=gscale(gbw);
ghf=dftfilt(f,H);
ghf=gscale(ghf);
ghe=histeq(ghf,256);

39
Nguyen Thanh Hai, PhD

University of Technology and Education

Faculty of Electrical & Electronic Engineering
Image Filtering
IMAGE FILTERING IN THE FREQUENCY DOMAIN
In the Ex 4.6, the gscale allows to normalize and change data type
containing the gray intensity in image. The gscale function is proposed:

function g=gscale(f)
fmin=min(f(:));
fmax=max(f(:));
for i=1:size(f,1)
for j=1:size(f,2)
g(i,j)=255*(f(i,j)-fmin)/(fmax-fmin);
end
end
g=uint8(g);
end

40
Nguyen Thanh Hai, PhD

20
 11/11/2015

University of Technology and Education

Faculty of Electrical & Electronic Engineering
Image Filtering

(a) Blurring satellite image (c) Image after the

emphasis highpass filter

(b) Image after the (d) Image after the histogram

Butterworth highpass filter equalization of image (c)
41
Nguyen Thanh Hai, PhD Ex 4.11. Satellite image after the highpass filter

University of Technology and Education

Faculty of Electrical & Electronic Engineering

Image Filtering
Problem: Calculate the convolution of the image A and the
mask h.
• Suppose the input image is

• And the kernel (mask)

42
Nguyen Thanh Hai, PhD

21
 11/11/2015

University of Technology and Education

Faculty of Electrical & Electronic Engineering

Image Filtering

The End

43
Nguyen Thanh Hai, PhD

22