Digital Image Processing Image Filtering

Table of Content
Digital Image Processing

Image Filtering
O. Le Meur
olemeur@irisa.fr
Univ. of Rennes 1
http://www.irisa.fr/temics/staff/lemeur/
January 2, 2011
1
Table of Content
1 Introduction
2 Point-to-point transformation
3 Linear ltering (neighborhood

operator)
4 Non Linear ltering
5 Conclusion
2
Introduction
Point-to-point transformation
Linear ltering (neighborhood operator)
Non Linear ltering
Conclusion
Introduction
1 Introduction

operator)
5 Conclusion
3
Introduction
Non Linear ltering
Conclusion
Introduction: image transformation
There exist 3 types of transformation:

Point to point transformation:
The output value at a specic coordinate
is dependent only on one input value but
not necessarily at the same coordinate;
Local to point transformation:

is dependent on the input values in the
neighborhood of that same coordinate;
Global to point transformation:

is dependent on all the values in the input
image.
Note that the complexity increases with the size of the considered neighborhood...
4
Introduction
Point-to-point transformation Spatial coordinates-based transformations
Linear ltering (neighborhood operator) Pixel values-based transformations
Non Linear ltering
Conclusion
1 Introduction
Spatial coordinates-based
transformations
Pixel values-based transformations
operator)
5 Conclusion
5
Introduction
Non Linear ltering
Conclusion
Spatial coordinates-based transformations
Remark:
This section is composed of several pictures extracted from
http://eeweb.poly.edu/~onur/lectures/lectures.html.
Let im[x , y ] be an input image of size N × N .
A spatial coordinates-based transformation, also called warping, aims at providing an
image IM [k , l ] from the input image im[x , y ]:
IM [k , l ] = im[x (k , l ), y (k , l )]
x (k , l ) and y (k , l ) are the transformations or the pixel warping functions. These functions
just modify the spatial coordinates of a pixel not its value;
Special cases to take into consideration:
the new coordinates [x (k , l ), y (k , l )] is out of the image IM . In this case,

IM [k , l ] = 0;
the new coordinates must be integers (Rounding operation, nearest integers...).
6
Introduction
Non Linear ltering
Conclusion
Transpose
The transpose tansformation is given by

x (k , l ) = l
y (k , l ) = k
7
Introduction
Non Linear ltering
Conclusion
Vertical Flip
The vertical ip tansformation is given by
x (k , l ) = N −k
y (k , l ) = l
The horizontal ip tansformation is given by

x (k , l ) = k
y (k , l ) = N −l
8
Introduction
Non Linear ltering
Conclusion
Translation
The translation tansformation is given by
x (k , l ) = k + Tk
y (k , l ) = l + Tl
where Tk and Tl are the transalation values for the x-axis and the y-axis respectively.
In the example below, we have (Tl = −50)
IM [k , l ] = im[x (k , l ), y (k , l )]
IM [k , l ] = im[k , l + Tl ]
Dierent transformations can be obtained depending on Tk and Tl values.

9
Introduction
Non Linear ltering
Conclusion
Rotation
The rotation tansformation is given by

x (k , l ) = (k − x0 )cos θ − (l − y0 )sinθ + x0
y (k , l ) = (k − x0 )sinθ + (l − y0 )cos θ + y0
where [x0 , y0 ] is the spatial coordinates of the center of the rotation and θ the angle.
Extracted from http://eeweb.poly.edu/ onur/lectures/lectures.html.
10
Introduction
Non Linear ltering
Conclusion
Wave
x (k , l ) = k x (k , l ) = k + α × sin(β × k )
y (k , l ) = l + α × sin(β × l ) y (k , l ) = l
α and β can be used to strengthen the eect.
11
Introduction
Non Linear ltering
Conclusion
Warp and swirl
The swirl eect is a rotation but the

x (k , l ) = k angle of the rotation θ varies with the
y (k , l ) sign(l − y0 ) × (l − y )2 + y pixel distance from the center of the
=
y0 0 0 image [x0 , y0 ]:
d (k − x0 )2 + (l − y0 )2
q
=
π
θ = r
512
If r −→ 0, θ is small...
12
Introduction
Non Linear ltering
Conclusion
Glass eect
A glass eect is obtained by adding a small and random displacement to each pixel:
x (k , l ) = k + (RAND (1, 1) − 1/2) × 10
y (k , l ) = l + (RAND (1, 1) − 1/2) × 10
Extracted from http://eeweb.poly.edu/ onur/lectures/lectures.html.
13
Introduction
Non Linear ltering
Conclusion
Summary
All 2D linear transformations:
x a b k

y =
c d l
Scale, rotation, mirror...
Properties:
Origin maps to origin;
Lines map to lines;
Ratios are preserved...
Ane transformations (linear transf. + translation):
x a b c k
    
y  = d e f   l 
w 0 0 1 w
Properties:
Origin does not necessarily map to origin;
Lines map to lines;
Ratios are preserved...
14
Introduction
Non Linear ltering
Conclusion
Summary
Ane transformations (linear transf. + translation):

Translation:
x 1 0 tx k
    
y  = 0 1 ty   l 
1 0 0 1 1
Scale:
x sx 0 0 k
    
y  = 0 sy 0  l 
1 0 0 1 1
2D in-plane rotation:
x cos θ −sinθ 0 k
    
y  =  sinθ cos θ 0  l 
1 0 0 1 1
15
Introduction
Non Linear ltering
Conclusion
Pixel values-based transformations
Let im[x , y ] be an input image of size N × N .
A pixel values-based transformation aims at providing an image IM [k , l ] from the

input image im[x , y ]:
IM [k , l ] = f (im[k , l ])
Noticed that the spatial coordinates of pixel are not modied. The function f is
used to modify the pixel values.
The simplest one is the identity function: f (p ) = p :

IM [k , l ] = f (im[k , l ])
IM [k , l ] = im[k , l ]
16
Introduction
Non Linear ltering
Conclusion
Histogram
An image histogram is a graphical representation of the tonal distribution in a digital
image. It plots the number of pixels for each tonal value.
Histogram gives information about the global distribution of an image.
Histogram plots the number of pixels in the image (vertical axis) with a particular
brightness value (horizontal axis).
(a) Original (b) Histogram (intensity)
17
Introduction
Non Linear ltering
Conclusion
Histogram
(a) Original (b) Histogram (intensity)
(c) Original (d) Histogram (intensity)
18
Introduction
Non Linear ltering
Conclusion
Histogram
High dynamic range in the last case

providing the best quality.
From R.C. Gonzalez, R.E. Woods, Digital image processing.
19
Introduction
Non Linear ltering
Conclusion
Histogram equalization
The goal is to increase the global contrast of images, especially when the usable data
of the image is represented by close contrast values.
Consider a discrete grayscale image {x } and let ni be the number of occurrences of
gray level i .
The probability of an occurrence of a pixel of level i in the image is
px (i ) = p(x = i ) = nni , 0 ≤ i < L, L being the total number of gray levels in the
image, n being the total number of pixels in the image.
Let us also dene the cumulative distribution function: cdfx (i ) = ij =0 px (j ).
P
We want to produce an image {y }, such that cdfy (i ) = iK .
y = cdfx (x ) × (max {x } − min {x }) + min {x }
20
Introduction
Non Linear ltering
Conclusion
Negative image
The negative image is obtained by f (p ) = 255 − p (pixel values are coded on 8 bits).
IM [k , l ] = f (im[k , l ])
IM [k , l ] = 255 − im[k , l ]
21
Introduction
Non Linear ltering
Conclusion
Piece continuous transformation
The objective of such transformation is:

to compress pre-determined ranges of values: Range compression;
to accentuate pre-determined ranges of values: Range stretching.
α1 × p 0 ≤ p < a1

α2 × (p − a1 ) a 1 ≤ p < a2




 ..




f (p ) = .
αi (p − ai −1 ) + i −1 α (α − α ai − 1 ≤ p < ai
P
1 j j j −1

j

 =
 ..




.

Obviously, we have:
αi < 1, compression;
αi > 1, stretching.
22
Introduction
Non Linear ltering
Conclusion
Example of contrast stretching:

Let f a function dened on three pieces:
 α1 p , 0 ≤ p < a1

f (p) =  α2 (p − a1 ) + α1 a1 , a1 ≤ p < a2
α3 (p − a2 ) + (α2 (a2 − a1 ) + α1 a1 ), a2 ≤ p ≤ 255
23
Introduction
Non Linear ltering
Conclusion
Example of contrast stretching:
Let imaging that α1 and α2 are null. The ltered image contains only grey level
belonging to [a1 , a2 ]. We just keep a slice of the image.
24
Introduction
Non Linear ltering
Conclusion
Binary thresholding:
0 im[x,y]<T

IM [x , y ] = 255 otherwise
Gamma correction:
IM [x , y ] = im[x , y ]γ
25
Introduction Denition
Point-to-point transformation Low-pass lters in spatial domain
Linear ltering (neighborhood operator) High-pass lters in spatial domain
Non Linear ltering Dierentiation lter in spatial domain
Conclusion Frequency domain ltering
Linear ltering with neighborhoods operator
1 Introduction

operator)
Denition
Low-pass lters in spatial domain
High-pass lters in spatial domain
Dierentiation lter in spatial
domain
Frequency domain ltering
5 Conclusion
26
Denition
Denition of neighborhood around a pixel of spatial coordinate (x , y ). The

neighborhood is called V (x , y ):
Two examples:
27
Denition
2D Finite Impluse Response (FIR) lter:

IM [x , y ] = (im ∗ h)[x , y ]
IM [x , y ] h(k , l ) im[x − k , y − l ]
X X
=
k ∈V (x ,y ) l ∈V (x ,y )
| {z } | {z } | {z }
Output Filter coe. Input
with h is the 2D impulse response called also the Point Spread Function (PSF) or the
kernel of the transform. It is composed of the lter coecients (nite length).
The gain of the lter is equal to
g= h(i , j )
X X
k ∈V (x ,y ) l ∈V (x ,y )
28
Denition
IM [x , y ] = h(k , l ) im[x − k , y − l ]
X X
k ∈V (x ,y ) l ∈V (x ,y )
| {z } | {z } | {z }
where H is the convolution kernel.

Example: the lter support is (3 × 5). The convolution kernel is:
h(−2, −1) h(−1, −1) h(0, −1) h(1, −1) h(2, −1)
 
H =  h(−2, 0) h(−1, 0) h(0, 0) h(1, 0) h(2, 0) 

h(−2, −1) h(−1, −1) h(0, −1) h(1, −1) h(2, −1)
29
Denition
Example for a neigborhood of size (2N + 1) × (2N + 1):

N=2
N N
IM [x , y ] = h(k , l )im[x −k , y −l ]
X X
k =−N l =−N
Number of multiplications per output point: O(N 2 )
30
Average lter
The most simple low-pass lter is the local averaging operation. The main eect of a
low-pass lter is a blurring. The size of the kernel is (2N + 1) × (2N + 1):
1 −N ≤ k , l ≤ − N
(
h (k , l ) = (2N +1)2
0 Otherwise
For N = 1, the convolution kernel is given by:
1 1 1
 
H = 1 1 1 1
9 1 1 1
31
Average lter
Three examples of averaging for dierent sizes of kernel. From the left-hand side to
the right-side, N = {1, 3, 8}:
The amount of blur increases with the size of the kernel (the number of operation too
O(N 2 )). In order to lter pixels located near the edges of the image, edge pixel values
are replicated to give sucient data (this is not the case here).
32
Gaussian lter
The kernel h is given by the following function
h(x , y ) = 1 2 exp x2 + y2

−
2πσ 2σ2
Each pixel's new value is set to a weighted average of that pixel's neighborhood. The
original pixel's value receives the heaviest weight (having the highest Gaussian value)
and neighboring pixels receive smaller weights as their distance to the original pixel
increases. This results in a blur that preserves boundaries and edges better than other,
more uniform blurring lters.
Note that the lter support is truncated...
33
Gaussian lter
h(x , y ) = 1 2 exp x2 + y2

−
2πσ 2σ2
Example of kernel:
σ = 0.84089642, N = 3:
Note that the center element (at [4, 4]) has the largest value, decreasing
symmetrically as distance from the center increases.
0.00000067 0.00002292 0.00019117 0.00038771 0.00019117 0.00002292 0.00000067
0.00002292 0.00078633 0.00655965 0.01330373 0.00655965 0.00078633 0.00002292
0.00019117 0.00655965 0.05472157 0.11098164 0.05472157 0.00655965 0.00019117

H = 0.00038771 0.01330373 0.11098164 0.22508352 0.11098164 0.01330373 0.00038771


0.00019117 0.00655965 0.05472157 0.11098164 0.05472157 0.00655965 0.00019117


0.00002292 0.00078633 0.00655965 0.01330373 0.00655965 0.00078633 0.00002292

0.00000067 0.00002292 0.00019117 0.00038771 0.00019117 0.00002292 0.00000067
Note that 0.22508352 (the central one) is 1177 times larger than 0.00019117
which is just outside 3σ.
34
Gaussian lter
h(x , y ) = 1 2 exp x2 + y2

−
2πσ 2σ2
Example of kernel:
σ = 0.6, N = 1:
1 2 1
 
H = 1 2 4 2
16 1 2 1
σ = 1, N = 2:
1 
9 18 9 1
1  9 81 162 81 9
H= 18

162 324 162 18
1444  9 81 162 81 9

1 9 18 9 1
35
Gaussian lter
Example of Gaussian ltering with σ = 2:
36
Gaussian lter
Increasing sigma increases the smoothing
37
Other low pass lters
2D Pyramidal lter:

1 2 3 2 1
2 4 6 4 2
H= 1 3

6 9 6 3
81 2 4 6 4 2

1 2 3 2 1
Conic lter: 
0 0 1 0 0
0 2 2 2 0
H= 1 1

2 5 2 1
25 0 2 2 2 0

0 0 1 0 0
38
High-pass lter
The high-pass ltered image can be thought of as the original image minus the low
pass ltered image.
N N
IM [x , y ] = im[x , y ] − h(k , l )im[x − k , y − l ]
X X
k =−N l =−N
N N
IM [x , y ] = h(k , l )im[x − k , y − l ]
X X
k =−N l =−N
with h the convolution kernel.
0 −1 0 −1 −1 −1 1 −2 1
     
H = −1 5 −1 H = −1 9 −1 H = −2 5 −2

0 −1 0 −1 −1 −1 1 −2 1
39
High-pass lter
Three examples of high-pass ltering for dierent sizes of kernel. From the left-hand
side to the right-side, N = {1, 3, 8}:
When the kernel's size increases, the ltering is more important and then the result is
less noisy.
40
Dierentiation lter
f 0 (x ) = ∆lim f (x + ∆x ) − f (x )
x →0 ∆x
Local variations of intensity are an important source of information in image
processing. These local variations are gradient (it measures the rate of change of the
function):
∂ im ∂ im

∇im[k , l ] = [k , l ], [k , l ]
∂x ∂y
In the illustration below: GX = ∂ im
∂x
[ k , l ] and GY y [k , l ].
= ∂∂im
41
Dierentiation lter
GX x [k , l ]:
= ∂∂im Kernel = hx = −1 1

∂ im
[k , l ] ≈ im[k + 1, l ] − im[k , l ]
∂x
GY y [k , l ]
= ∂∂im Kernel = hy = −1 1 T

∂ im
[k , l ] ≈ im[k , l + 1] − im[k , l ]
∂y
However, most of the time, we use the following kernel −1 0
1 and

T
−1 0 1 (phase =zero). Below, from the left-hand side to the right: original,

−1 1 , −1 0 1 .

42
Dierentiation lter
However, these lters are very sensitive to the noise. In order to enhance the
robustness, these lters are combined with a blurring lter:
−1 0 1
 
hx =  −2 0 2
−1 0 1
−1 −2 −1
 
hy =  0 0 0
1 2 1
These kernel are the Sobel's kernel. The blurring lter is 1 2 1.

IMx [k , l ] = (im ∗ hx )[k , l ]

IMy [k , l ] = (im ∗ hy )[k , l ]
In the same vein, Prewitt and Robert's lters.
43
Dierentiation lter
Norm of the gradient: k∇IM [k , l ]k2 = IMx [k , l ]2 + IMy [k , l ]2 ;

p
IMy [k ,l ]
Its orientation: arg (∇IM [k , l ]) = arctan

IMx [k ,l ]
From left-hand side to the right: IMx , IMy and the norm.
44
Dierentiation lter
The Laplacian of a picture is the second derivative:
∂ 2 im
[k , l ] ≈ im[k + 1, l ] + im[k − 1, l ] − 2 × im[k , l ]
∂2x
∂ 2 im
[k , l ] ≈ im[k , l + 1] + im[k , l − 1] − 2 × im[k , l ]
∂2y
∂ 2 im ∂ 2 im
∇2 im[k , l ] = [k , l ] + 2 [k , l ]
∂ x2 ∂ y
∇2 im[k , l ]
≈ im[k + 1, l ] + im[k − 1, l ] + im[k , l + 1] + im[k , l − 1] − 4 × im[k , l ]
For a 4-neighborhood, the kernel is given by
0 1 0
 
h = 1 −4 1
0 1 0
We can extent this kernel to compute the laplacian in all directions (8-neighborhood):
1 1 1
 
h = 1 −8 1
1 1 1
45
Dierentiation lter
The second order derivatives have a stronger response to ne details (e.g. thin lines)
than the rst order derivatives.
46
where H is the convolution kernel.
IM [x , y ] = (im ∗ h)[x , y ]
im1 [x , y ] ∗ im2 [x , y ] →
F
IM1 [u, v ] × IM2 [u, v ]
im1 [x , y ] × im2 [x , y ] →
F
IM1 [u, v ] ∗ IM2 [u, v ]
When the size of the kernel is large, it is better to apply the lter in the frequency domain.
For more information:

Digital Image Processing, by R. C. Gonzalez and R. E. Woods, 3rd edition, Pearson Prentice
Hall, 2008.
47
We can spatially lter an image by Fourier transforming and applying a frequency
lter:
IM [x , y ] = im[x , y ] ∗ h[x , y ]
IM
˜ [u , v ] = IM [u, v ] × H [u, v ]
where, H [u , v ] is the lter in the frequency function.
48
Ideal low pass lter
From the left-hand side to the right: Ideal low pass lter transfert function, lter
displayed as an image, lter radial cross section.
1 D (u, v ) ≤ D0

H (u, v ) =
0 D (u, v ) > D0
With D the euclidean distance from the spectrum center ( N2 , N2 ).
49
Ideal low pass lter
Low pass ltering:
Ringing and blurring
50
Butterworth low pass lter
1
H (u , v ) =
D (u ,v ) 2n
1+

D0
51
Butterworth low pass lter
Top: spatial representation of the lter for dierent orders;

Bottom: intensity proles through the center of the lters.
Butterworth low pass ltering:
Smooth transition in blurring, no ringing is present.
52
Gaussian low pass lter
D 2 (u, v )

H (u, v ) = exp −
2D02
with D0 = σ. Gaussian low pass ltering:
Smooth transition in blurring, no ringing is present.

53
Low pass lter
Comparison between the ideal, the Butterworth and the gaussian low pass lter:
Ideal
Butterworth
Gaussian
54
High pass ltering in the frequency domain
HHP (u, v ) = 1 − HLP (u , v )
55
High pass ltering in the frequency domain
Ideal high-pass lters enhance edges but suer from ringing artefacts, just like
Ideal LPF;
smoother results with the two others.
56
Example of two band-pass lters:
57
Laplacian in the frequency domain
We remind (see previous lecture):
∂ 2 im ∂ 2 im
∇2 im[k , l ] = [k , l ] + 2 [k , l ]
∂2x ∂ y
d n x (t ) F
dt n −→ (j 2π f ) X (f ).
n
From this, it follows that

∂ 2 im ∂ 2 im
[k , l ] + 2 [k , l ] −→ −(u 2 + v 2 )IM [u , v ]
F
∂ x2 ∂ y
The Laplacian lter is then implemented in the frequency domain by

H (u, v ) = −(u 2 + v 2 )
58
Laplacian in the frequency domain
Finally, to compute the Laplacian, we need :

1 to compute the Fourier transform of the picture;
2 to multiply the spectrum by −(u 2 + v 2 );

3 to compute the inverse Fourier transform.
59
The cortex transform
The cortex transform is rst described by A. Watson as the modeling of the neural
response of retinal cells to visual stimuli.
The cortex lter in the frequency domain is
cortex[bi ,θi ] (ρ, θ) = dombi × fanθi (θ)
Where,
bi and θi represent the frequency band and the index of orientation, respectively;
(ρ, θ) are polar coordinates.
The cortex transform decomposes the input image im[x , y ] into a set of subband
images B[bi ,θi ] [k , l ]:
B[bi ,θi ] [k , l ] F −1 cortex[bi ,θi ] (ρ, θ) × F {im[x , y ]}

=
60
The cortex transform
Frequency responses of several cortex lters (brightness represents gain for the given
spatial frequency).
Complete set of cortex lters
61
Point-to-point transformation Rank ltering
Linear ltering (neighborhood operator) Homomorphic ltering
Non Linear ltering Adaptive ltering
Conclusion
Introduction
1 Introduction

operator)

Denition
Rank ltering
Homomorphic ltering
Adaptive ltering
Conditional mean
Anisotropic Kuwahara ltering
Bilateral ltering
5 Conclusion
62
Conclusion
Denition
The most important drawback of the linear ltering is that all pixels in the image are
modied by the ltering process.
To overcome this problem, non linear ltering is used. It aims, for instance, to protect
some parts of the picture having particular features (edges...) or to remove data
without blurring the whole image (impluse noise).
Three kinds of lters:

Rank ltering;
Adaptive linear ltering;
Morphological operators.
63
Conclusion
Rank ltering
Rank lters are based on three steps:
1 Data selection, also called

windowing;
2 Data ranking (in ascendant order);
3 1D linear weighting of the ordered
data list.
64
Conclusion
Special case of a generalized rank lter
If all weights of the linear lter are null, except one in the median position. This lter
is called a median lter.
IM [x , y ] = MED (im[xi , yi ]|[xi , yi ] ∈ V [x , y ])
V [x , y ] is the neighborhood (a set of N samples).
if the size of the neighborhood is odd, the output value is the median value;
if the size is even, the output value is the average of the two middles values.
The median lter is very ecient in ltering signals corrupted by impulsive noise but it
is not very ecient in gaussian noise environment.
65
Conclusion
Special case of a generalized rank lter
However, when the number of the samples is large, the ordering procedure becomes
cumbersome.
Idea: the median lter is taken over the outputs of several FIR substructures and the
number of the substructures is much smaller than the number of the data samples
inside the lter window.
IM [x , y ] = MED (y (1), . . . , y (m))
where, m is linear FIR lters.
66
Conclusion
Homomorphic ltering
Homomorphic ltering is a generalized technique for signal and image processing, in-
volving a nonlinear mapping to a dierent domain in which linear lter techniques are
applied, followed by mapping back to the original domain.
In many case, we want to remove shading eects from an image. The objective is then
to enhance high frequencies;
to attenuate low frequencies (but ne details have to be preserved).
Consider the following model of image formation:
im[x , y ] = i (x , y ) × r (x , y )
| {z } | {z }
illumination reection
The illumination component varies slowly and then aects low frequencies mostly;
The reection component varies faster and then aects the high frequencies
mostly.
67
Conclusion
Homomorphic ltering
What is the solution to separate LF and HF?
im[x , y ] = i [x , y ] × r [x , y ]
Fourier transform
IM [u, v ] = I [u , v ] ∗ R [u , v ]
Due to the convolution in the frequency domain, LF and HF from i [x , y ] and r [x , y ]

are mixed together...
Too complex to lter LF and HF in such condtion.
68
Conclusion
Homomorphic ltering
What is the solution to separate LF and HF? We can take the log !
im[x , y ] = i [x , y ] × r [x , y ]
log (im[x , y ]) = log (i [x , y ] × r [x , y ])
log (im[x , y ]) = log (i [x , y ]) + log (r [x , y ])
1 Take the log and apply the Fourier tansform to the new signal:
log (im[x , y ]) = log (i [x , y ]) + log (r [x , y ])
F (log (im[x , y ])) = F (log (i [x , y ])) + F (log (r [x , y ]))
Z [u , v ] = Ilog [u, v ] + Rlog [u, v ]
2 Filtering in the frequency domain: H [u , v ]
Z [u , v ] × H [u , v ] = Ilog [u, v ] × H [u, v ] + Rlog [u, v ] × H [u, v ]
3 Take the inverse Fourier transform and apply the exponential function:
(Z [u , v ] × H [u , v ]) = F Ilog [u, v ] × H [u, v ] + F −1 Rlog [u, v ] × H [u, v ]

−1 −1
F
z [x , y ] = i [x , y ] + er [x , y ]
e
exp (z [x , y ]) exp ei [x , y ] + exp (er [x , y ])

=
69
Conclusion
Homomorphic ltering
γH
H [u , v ] = γL + 2n
1 + √ D20
u +v 2
with,
γL a parameter that aects low frequencies;
γH a parameter that aects high frequencies;
70
Conclusion
Adaptive ltering
An adaptive lter is a lter that self-adjusts its transfer function according to an

optimizing algorithm.
The goal is still to smooth the signal. However, we want to preserve edges...
Filtering by pixel grouping;

Conditional mean, Bilateral ltering and mean shift lter;
Diusion.
71
Conclusion
Conditional mean
IM [x , y ] h(k , l ) im[x − k , y − l ]
X X
=
k ∈V (x ,y ) l ∈V (x ,y )
| {z } | {z } | {z }
Principle: pixels in a neighbourhood are averaged only if they dier from the central
pixel by less than a given threshold.
1 if |IM [x − k , y − k ] − IM [k , l ]| < TH

h(k , l ) =
0 Otherwise.
Example with a neighbourhood equal (2 × 3 + 1)(2 × 3 + 1),TH = 32:
72
Conclusion
Method proposed by Kuwahara and adapted by Nagao in 1980.

Example for a window 5 × 5:
Selection of the sub-domain that

has the minimum variance (9
windows for Nagao);
Replace the value of the central
pixel by the average value of the
sub-domain having the minimum
variance.
73
Conclusion
74
Conclusion
Bilateral ltering
The idea is to use a weighted ltering but with an outlier rejection.
Pixels that are very dierent in intensity from the central pixel are weighted less even
though they may be in close proximity to the central pixel.
This is applied as two Gaussian lters at a localized pixel neighborhood:
1
IM [x , y ] im[k , l ] c[x ,y ] [k , l ]s[x ,y ] [k , l ]
X
=
k ,l c[x ,y ] [k , l ]s[x ,y ] [k , l ]
P
k ,l | {z }
h[k ,l ]
One in the spatial domain named the domain lter:

d ([x , y ], [k , l ])

c[x ,y ] [k , l ] = exp −
2σd
where d is Euclidean distance.
One in the intensity space named the range lter:
φ(im[x , y ], im[k , l ])

s[x ,y ] [k , l ] = exp −
2σr
where φ is a suitable measure of distance in intensity space.
75
Conclusion
Bilateral ltering
76
Conclusion
Bilateral ltering
Low contrast texture has been removed and edges are well preserved.
77
Introduction
Non Linear ltering
Conclusion
Conclusion
1 Introduction

operator)
5 Conclusion
78
Introduction
Non Linear ltering
Conclusion
Conclusion
Image Transformation: global to point; point to point; local to point;

Histogram;
Linear Filtering;
Non linear ltering.
79

Digital Image Processing Image Filtering

Uploaded by

Document Information

Original Description:

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Digital Image Processing Image Filtering

Uploaded by

Copyright:

Available Formats

Table of Content

Digital Image Processing

3 Linear ltering (neighborhood

4 Non Linear ltering

3 Linear ltering (neighborhood

4 Non Linear ltering

There exist 3 types of transformation:

Local to point transformation:

Global to point transformation:

4 Non Linear ltering

Special cases to take into consideration:

the new coordinates [x (k , l ), y (k , l )] is out of the image IM . In this case,

The transpose tansformation is given by

The horizontal ip tansformation is given by

Dierent transformations can be obtained depending on Tk and Tl values.

The rotation tansformation is given by

Extracted from http://eeweb.poly.edu/ onur/lectures/lectures.html.

α and β can be used to strengthen the eect.

The swirl eect is a rotation but the

Extracted from http://eeweb.poly.edu/ onur/lectures/lectures.html.

Ane transformations (linear transf. + translation):

Let im[x , y ] be an input image of size N × N .

A pixel values-based transformation aims at providing an image IM [k , l ] from the

The simplest one is the identity function: f (p ) = p :

(a) Original (b) Histogram (intensity)

(a) Original (b) Histogram (intensity)

(c) Original (d) Histogram (intensity)

High dynamic range in the last case

From R.C. Gonzalez, R.E. Woods, Digital image processing.

We want to produce an image {y }, such that cdfy (i ) = iK .

y = cdfx (x ) × (max {x } − min {x }) + min {x }

The objective of such transformation is:

Example of contrast stretching:

3 Linear ltering (neighborhood

4 Non Linear ltering

Denition of neighborhood around a pixel of spatial coordinate (x , y ). The

2D Finite Impluse Response (FIR) lter:

where H is the convolution kernel.

H =  h(−2, 0) h(−1, 0) h(0, 0) h(1, 0) h(2, 0) 

Example for a neigborhood of size (2N + 1) × (2N + 1):

Number of multiplications per output point: O(N 2 )

0.00000067 0.00002292 0.00019117 0.00038771 0.00019117 0.00002292 0.00000067

Example of Gaussian ltering with σ = 2:

Increasing sigma increases the smoothing

H = −1 5 −1 H = −1 9 −1 H = −2 5 −2

These kernel are the Sobel's kernel. The blurring lter is 1 2 1.

IMx [k , l ] = (im ∗ hx )[k , l ]

In the same vein, Prewitt and Robert's lters.

Norm of the gradient: k∇IM [k , l ]k2 = IMx [k , l ]2 + IMy [k , l ]2 ;

where H is the convolution kernel.

For more information:

where, H [u , v ] is the lter in the frequency function.

Low pass ltering:

Ringing and blurring

Top: spatial representation of the lter for dierent orders;

Smooth transition in blurring, no ringing is present.

Smooth transition in blurring, no ringing is present.

HHP (u, v ) = 1 − HLP (u , v )

Example of two band-pass lters:

From this, it follows that

The Laplacian lter is then implemented in the frequency domain by

Finally, to compute the Laplacian, we need :

3 Linear ltering (neighborhood

4 Non Linear ltering

3 Linear ltering (neighborhood

4 Non Linear ltering

4 Non Linear ltering

The horizontal ip tansformation is given by

Dierent transformations can be obtained depending on Tk and Tl values.

α and β can be used to strengthen the eect.

The swirl eect is a rotation but the

Ane transformations (linear transf. + translation):

3 Linear ltering (neighborhood

4 Non Linear ltering

Denition of neighborhood around a pixel of spatial coordinate (x , y ). The

2D Finite Impluse Response (FIR) lter:

Example of Gaussian ltering with σ = 2:

These kernel are the Sobel's kernel. The blurring lter is 1 2 1.

In the same vein, Prewitt and Robert's lters.

where, H [u , v ] is the lter in the frequency function.

Low pass ltering:

Top: spatial representation of the lter for dierent orders;

Example of two band-pass lters:

The Laplacian lter is then implemented in the frequency domain by

Complete set of cortex lters

3 Linear ltering (neighborhood

4 Non Linear ltering

Three kinds of lters:

Rank lters are based on three steps:

An adaptive lter is a lter that self-adjusts its transfer function according to an

One in the spatial domain named the domain lter:

3 Linear ltering (neighborhood

4 Non Linear ltering