Linear Systems

• Many image processing (filtering)

operations are modeled as a linear system

δ(x,y) Linear System h(x,y)

g ( x , y )  f ( x, y )  h ( x , y ) 



f ( x, y)h( x  x, y  y)dxdy

E.G.M. Petrakis Filtering 1

 Impulse Response
• System’s output to an impulse δ(x,y)

δ(x,y)

0 t

E.G.M. Petrakis Filtering 2

 Space Invariance
• g(x,y) remains the same irrespective of the
position of the input pulse
δ(x-x0,y-y0) Space Inv. Syst. h(x-x0,y-y0)

• Linear Space Invariance (LSI)

af1(x,y)+bf2(x,y) LSI System ah1(x,y)+bh2(x,y)

E.G.M. Petrakis Filtering 3

 Discrete Convolution
• The filtered image is described by a discrete
convolution
g (i, j )  f (i, j )  h(i, j ) 
n m

  f (k , l )h(i  k , j  l )
k 1 l 1

• The filter is described by a n x m discrete

convolution mask
E.G.M. Petrakis Filtering 4
 Computing Convolution
• Invert the mask g(i,j) by 180o
– not necessary for symmetric masks
• Put the mask over each pixel of f(i,j)
• For each (i,j) on image
h(i,j)=Ap1+Bp2+Cp3+Dp4+Ep5+Fp6+Gp7+Hp8+Ip9

E.G.M. Petrakis Filtering 5

 Image Filtering
• Images are often corrupted by random
variations in intensity, illumination, or have
poor contrast and can’t be used directly
• Filtering: transform pixel intensity values to
reveal certain image characteristics
– Enhancement: improves contrast
– Smoothing: remove noises
– Template matching: detects known patterns

E.G.M. Petrakis Filtering 6

 Template Matching
• Locate the template in the image

E.G.M. Petrakis Filtering 7

 Computing Template Matching
• Match template with image at every pixel
– distance  0 : the template matches the image
at the current location
m n
D x, y     f x, y  t x  x, y  y 
2 2

x  0 y   0

– t(x,y): template
– M,N size of the template
E.G.M. Petrakis Filtering 8
 D  x, y  
2

M N

  f x, y  t x  x, y  y 

x 1 y 1
2


M N

 f x, y 2 background


x 1 y 1
M N
constant

 t x  x, y  y 
x 1 y 1
2
 correlation:
convolution of
M N f(x,y) with t(-x,-y)
2 f  x, y t  x  x, y   y 
x
E.G.M. Petrakis
1 y 1 Filtering 9
 true best match

template image correlation

1 1 1 1 1 0 0 0 4 5 3  
1 1 1 1 1 1 0 0 5 7 3  
1 1 1 1 0 1 0 0 3 5 8 7 
0 0 0 5 0  6 6  
0 0 0 0 0     

noise false match

E.G.M. Petrakis Filtering 10
 Observations
• If the size of f(x,y) is n x n and the size
of the template is m x m the result is
accumulated in a (n-m-1) x (n+m-1) matrix
• Best match: maximum value in the
correlation matrix but,
– false matches due to noise

E.G.M. Petrakis Filtering 11

 Disadvantages of Correlation
• Sensitive to noise
• Sensitive to variations in orientation and scale
• Sensitive to non-uniform illumination
• Normalized Correlation (1:image, 2:template):
E q1q2   E q1 E q2 
N  x, y  
 q1  q2 

 q  E q   E q
2 2

1
2

– E : expected value

E.G.M. Petrakis Filtering 12

 Histogram Modification
• Images with poor contrast usually contain
unevenly distributed gray values
• Histogram Equalization is a method for
stretching the contrast by uniformly
distributing the gray values
– enhances the quality of an image
– useful when the image is intended for viewing
– not always useful for image processing

E.G.M. Petrakis Filtering 13

 Example
• The original image has very poor contrast
– the gray values are in a very small range
• The histogram equalized image has better contrast

E.G.M. Petrakis Filtering 14

 Histogram Equalization Methods
• Background Subtraction: subtract the
“background” if it hides useful information
– f’(x,y) = f(x,y) – fb(x,y)
• Static & Dynamic histogram equalization
methods
– Histogram scaling (static)
– Statistical scaling (dynamic)

E.G.M. Petrakis Filtering 15

 Static Histogram Scaling
• Scale uniformly entire histogram range:
– [z1,zk]: available range of gray values:
– [a,b]: range of intensity values in image:
– scale [a,b] to cover the entire range [z1,zk]
– for each z in [a,b] compute
zk  z1
z'  ( z  a )  z1
ba
– the resulting histogram may have gaps
E.G.M. Petrakis Filtering 16
 Statistical Histogram Scaling
• Fills all histogram bins continuously
– pi : number of pixels at level zi input histogram
– qi : number of pixels at level zi output histogram
– k1= k2=… : desired number of pixels in histogram bin
• Algorithm:
1. Scan the input histogram from left to right to find k1:
k1 1 k1

p
i 1
i  q1   pi
i 1
– all pixels with values z1,z2,…,zk-1 become z1
E.G.M. Petrakis Filtering 17
 Algorithm (conted)
2. Scan the input histogram from k1 and to the right to
find k2:
k2 1 k2

p
i 1
i  q1  q2   pi
i 1

– all pixels zk1,zk1+1,…,zk2 become z2

• Continue until the input histogram is exhausted
– might also leave gaps in the histogram

E.G.M. Petrakis Filtering 18

 Noise
• Images are corrupted by random variations
in intensity values called noise due to non-
perfect camera acquisition or environmental
conditions.
• Assumptions:
– Additive noise: a random value is added at each
pixel
– White noise: The value at a point is
independent on the value at any other point.
E.G.M. Petrakis Filtering 19
 Common Types of Noise
 Salt and pepper noise: random occurrences of both
black and white intensity values
 Impulse noise: random occurrences of white
intensity values
 Gaussian noise: impulse noise but its intensity
values are drawn from a Gaussian distribution
 noise intensity value: 2
 k2
 k: random value in [a,b]
g ( x, y )  e 2
 σ : width of Gaussian
 models sensor noise (due to camera electronics)

E.G.M. Petrakis Filtering 20

 Examples of Noisy Images

a. Original image
b. Original image
c. Salt and pepper noise
d. Impulse noise
e. Gaussian noise
E.G.M. Petrakis Filtering 21
 Noise Filtering
• Basic Idea: replace each pixel intensity value with
an new value taken over a neighborhood of fixed
size
– Mean filter
– Median filter
• The size of the filter controls degree of smoothing
– large filter  large neighborhood  intensive
smoothing

E.G.M. Petrakis Filtering 22

 Mean Filter
• Take the average of intensity values in a m x
n region of each pixel (usually m = n)
– take the average as the new pixel value

1
hi, j    f k , l 
mn km ln
– the normalization factor mn preserves the range
of values of the original image
E.G.M. Petrakis Filtering 23
 Mean Filtering as Convolution
• Compute the convolution of the original
image with
1 1 1
1  
g i, j   1 1 1
3 3  
1 1 1

– simple filter, the same for all types of noise

– disadvantage: blurs image, detail is lost
E.G.M. Petrakis Filtering 24
 Size of Filter
• The size of the filter controls the amount of
filtering (and blurring).
– 5 x 5, 7 x 7 etc. 1 1 1 1 1
1 1 1 1 1
1  
g (i , j )  1 1 1 1 1
55  
1 1 1 1 1

1 1 1 1 1


– different weights might also be used

– normalize by sum of weights in filter
E.G.M. Petrakis Filtering 25
 Examples of Smoothing

• From left to right: results of 3 x 3, 5 x 5 and 7 x 7 mean filters

E.G.M. Petrakis Filtering 26

 Median Filter
• Replace each pixel value with the median
of the gray values in the region of the
pixel:
1. take a 3 x 3 (or 5 x 5 etc.) region centered
around pixel (i,j)
2. sort the intensity values of the pixels in the
region into ascending order
3. select the middle value as the new value of
pixel (i,j)
E.G.M. Petrakis Filtering 27
 Computation of Median Values

• Very effective in removing salt and pepper or impulsive

noise while preserving image detail
• Disadvantages: computational complexity, non linear filter
E.G.M. Petrakis Filtering 28
 Examples of Median Filtering

• From left to right: the results of a 3 x 3, 5 x 5 and 7 x 7 median filter

E.G.M. Petrakis Filtering 29

 Gaussian Filter
• Filtering with a m x m mask
– the weights are computed according to a
Gaussian function: i2  j2
 2
g (i, j )  c  e 2
– σ is user defined
1 4 7 10 7 4 1
4 12 26 33 26 12 4 

Example: 7 26 55 71 55 26 7
 
m=n=7 10 33 71 91 71 33 10
7 26 55 71 55 26 7
σ2 = 2  
4 12 26 33 26 12 4
1 1 
 4 7 10 7 4
E.G.M. Petrakis Filtering 30
 Properties of Gaussian Filtering
• Gaussian smoothing is very effective for removing
Gaussian noise
• The weights give higher significance to pixels near the
edge (reduces edge blurring)
• They are linear low pass filters
• Computationally efficient (large filters are implemented
using small 1D filters)
• Rotationally symmetric (perform the same in all directions)
• The degree of smoothing is controlled by σ (larger σ for
more intensive smoothing)

E.G.M. Petrakis Filtering 31

 Gaussian Mask

• A 3-D plot of a 7 x & Gaussian mask: filter symmetric and isotropic

E.G.M. Petrakis Filtering 32

 Gaussian Smoothing

• The results of smoothing an image corrupted with

Gaussian noise with a 7 x 7 Gaussian mask
E.G.M. Petrakis Filtering 33
 Computational Efficiency
• Filtering twice with g(x) is equivalent to
filtering with a larger filter with    2
• Assumptions x2

g x, y   e 2 2

h  x, y   f  x , y   g  x , y 
hx, y   f  x, y   g  x, y   g  x, y 
E.G.M. Petrakis Filtering 34
  2
 2 
 x  2

g x   g x    e 2
e 2 2
d 

2 2
x  x 
     
  2  2 
 
 e

2 2
e 2 2
d 

 2 x2 
 2  
   2  x2
 
d   e  2


e
2
 2 2 2



E.G.M. Petrakis Filtering 35

 Observations
• Filter an image with a large Gaussian 2
– equivalently, filter the image twice with a
Gaussian with small σ
– filtering twice with a m x n Gaussian is
equivalent to filtering with a (n + m - 1) x (n +
m - 1) filter
– this implies a significant reduction in
computations

E.G.M. Petrakis Filtering 36

 Gaussian Separability
h i, j   f i, j   g i, j  
m n
  g k , l  f i  k , j  l  
k 1 l 1

m n 
k 2
l 2 
  e 2 2
f i  k , j  l  
k 1 l 1

E.G.M. Petrakis Filtering 37

 k
 2
l 2

m  2  n  2 
 e 2
 e
2
f i  k , j  l  
k 1
l 1
 
 h' 
m k2
 2
 e 2
hi  k , j  1-D Gaussian
k 1 horizontally

1-D Gaussian
vertically
• The order of convolutions can be reversed
E.G.M. Petrakis Filtering 38
• An example of the separability of Gaussian
convolution
– left: convolution with vertical mask
– right: convolution with horizontal mask
E.G.M. Petrakis Filtering 39
 Gaussian Separability
• Filtering with a 2D Gaussian can be
implemented using two 1D Gaussian
horizontal filters as follows:
– first filter with an 1D Gaussian
– take the transpose of the result
– convolve again with the same filter
– transpose the result
• Filtering with two 1D Gausians is faster !!
E.G.M. Petrakis Filtering 40
 a. Noisy image
b. Convolution with
1D horizontal
mask
c. Transposition
d. Convolution with
same 1D mask
e. Transposition 
smoothed image

E.G.M. Petrakis Filtering 41