TDI2131 Digital Image Processing

Image Restoration & Noise

Removal
Lecture 7

John See
Faculty of Information Technology
Multimedia University

Some portions of content adapted CYPee's notes, MMU. Most figures from Gonzalez/Woods 1
 Lecture Outline
● Degradation & Restoration Process Models
● Noise Models
● Restoration in the presence of noise only (spatial
filtering)
● Restoration of Periodic Noise
● Restoration in the presence of degradation & noise
(frequency domain filtering)

2
 Some Announcements
● Reminder: Assignment 1 is due this Friday.

3
 Image Restoration
● Image Restoration – To improve the appearance of an image
by application of a restoration process that uses a
mathematical model for image degradation
● Types of degradation:
– Blurring caused by motion or atmospheric disturbance
– Geometric distortion caused by imperfect lenses
– Superimposed interference pattern caused by mechanical systems
– Noise from electronic sources

4
How are these images restored?

5
Image Restoration: The Idea

Example
degraded
images Develop Develop
degradation inverse degradation
model process
Knowledge of
image creation
process

Apply Output
Input
inverse degradation image
image
process fˆ ( x, y )
g(x,y)

6
Degradation/Restoration Process Model

● Consists of 2 parts – Degradation function & Noise function

● General model in spatial domain:
g ( x, y ) = h ( x, y ) * f ( x, y ) + n ( x , y )
– g(x,y): degraded image, h(x,y): degradation function,
f(x,y): original image, n(x,y): additive noise function
7
 Degradation Model in Freq. Domain

● General model in frequency domain:

G (u, v) = H (u , v) F (u, v) + N (u, v)
– G(x,y): Fourier transform of degraded image,
H(x,y): Fourier transform of degradation function,
F(x,y): Fourier transform of original image,
N(x,y): Fourier transfor of additive noise function
● What needs to be done??
FIND Degradation function and Noise model

8
 Noise Models
● Noise – Any undesired information that contaminates
an image
● Variety of sources:
– Digital image acquisition process, e.g. For CCD (charged
coupled device) camera, electronics signal fluctuations in
detector, caused by thermal energy and light levels
– During image transmission, e.g. Wireless network, may be
corrupted by lightning or other atmospheric disturbance

9
 Noise Models
● Assumptions of Noise models in this course:
– Noise is independent of spatial coordinates (except for
spatially periodic noise)
– Noise is uncorrelated w.r.t the image (pixel values)
●
Spatial Noise Descriptor – concern of the statistical behavior
of the intensity values in the noise component of the model
– Characterized by probability density function (PDF)

10
 Noise Models
● Typical image noise models that can be modeled:
– Gaussian
– Rayleigh
– Erlang (Gamma)
– Exponential
– Uniform
– Impulse (salt-and-pepper)

11
 Noise Models - PDFs
● Gaussian Noise
1 − ( z − µ ) 2 / 2σ 2
p( z ) = e
2πσ 2

where
z: gray scale
µ = mean (average)
σ = standard deviation

12
 Noise Models - PDFs
● Rayleigh Noise
2
 ( z − a ) exp(−( z − a ) / b for z ≥ a
2
p( z ) =  b
0 for z < a

where
mean = a + πb/ 4
b(4 − π )
variance = 4

13
 Noise Models - PDFs
● Uniform Noise
 1
 for a ≤ z ≤ b
p ( z ) =  (b − a )
0
 elsewhere

where
mean = (a + b) / 2
variance = (b − a ) 2 /12

14
 Noise Models - PDFs
● Uniform Noise
 Pa for z = a (pepper)

p( z ) =  Pb for z = b (salt)
 0 otherwise

where
b>a

15
Example: Noise Models

16
Example: Noise Models

17
Restoration in Presence of Noise Only –
Spatial Filtering
● Consider an image degraded with only additive noise.
The degradation model is further simplified as
g ( x , y ) = f ( x, y ) + n ( x , y )

in spatial domain, and

G (u , v) = F (u , v) + N (u , v)

in frequency domain

18
 Noise Removal with Spatial Filters
● Spatial filters can effectively remove various types of
noise in digital images
● Typically operate on small neighborhoods, from 3x3
to 11x11.
● Some can be implemented as convolution masks

19
 Mean Filters
● Arithmetic Mean Filter
– Computes the average value of the corrupted image g(x,y)
in area defined by Sxy. Sxy represents the set of
coordinates in a rectangular subimage window of size
mxn, centred at point (x,y)
1
fˆ ( x, y ) = ∑
mn ( s , t )∈S xy
g ( s, t )

– The operation can be implemented using convolution

mask
– For random noise

20
 Mean Filters
● Geometric Mean Filter
– The image restored using a geometric mean filter:
1
  mn
fˆ ( x, y ) =  ∏ g ( s, t ) 
 ( s , t )∈Sxy 
– For random noise: lose less detail than Arithmetic Mean
● Harmonic Mean Filter
– The image restored using a harmonic mean filter:
mn
fˆ ( x, y ) =
1
∑
( s , t )∈S xy g ( s, t )
– For salt noise

21
 Mean Filters
● Contraharmonic Mean Filter
– The image restored with a contraharmonic mean filter:
∑
( s , t )∈S xy
g ( s, t )Q +1
fˆ ( x, y ) =
∑
( s , t )∈S xy
g ( s , t )Q

where Q is called the order of filter

– Positive Q: pepper
– Negative Q: salt
– Q=0: Arithmetic mean
– Q=-1: Harmonic mean
22
Example: Mean Filters

23
Example: Mean Filters (Cont'd)

24
Example: Mean Filters (Cont'd)

25
 Order-Statistics Filters
● Order Filter – based on a specific type of image statistics
called order statistics, sometimes known as Order-Statistics
Filter
● Order Statistics: Arranges all the pixels in sequential order
(smallest to largest), based on gray-level value
● The selection of the value to be replaced in the center pixel,
is determined by the statistical function used (min, max,
median, etc.)

26
 Max & Min Filters
● Minimum Filter – select the smallest value within an ordered
window of pixel values, denoted as
fˆ ( x, y ) = min {g ( s, t )}
( s , t )∈S xy

– Works best when the noise is primarily of the salt-type

(high value)
● Maximum Filter – select the largest value within an ordered
window of pixel values, denoted as
fˆ ( x, y ) = max {g ( s, t )}
( s , t )∈S xy

– Works best for pepper-type noise (low value)

27
 Median Filters
● Median Filter – select the middle pixel value within an
ordered window of pixel values, denoted as
fˆ ( x, y ) = median{g ( s, t )}
( s ,t )∈S xy

– Works best with salt-and-pepper noise (both high and low

values)
● Better noise remover than Averaging Filter, which causes
blurry edges and details in image, thus not effective against
impulse (salt-and-pepper) noise
● Preserve line structures

28
Example: Median Filter

29
Example: Min & Max Filters

30
 Midpoint Filter
● Midpoint Filter – select the average of the maximum and
minimum pixel values within the window, denoted as
ˆf ( x, y ) = 1  max {g ( s, t )} + min {g ( s, t )}
2  ( s , t )∈S xy ( s , t )∈S xy 
● Useful for Gaussian and uniform noise

31
 Alpha-trimmed Mean Filter
● Alpha-trimmed Mean Filter – select the average of the values
within the window, but with some of the endpoint-ranked
values excluded
● Suppose we delete d/2 lowest and d/2 highest gray level
values of g(s,t) in the neighborhood Sxy. Let gr(s,t) represent
the remaining mn-d pixels, the remaining pixels are
averaged: 1
fˆ ( x, y ) =
mn − d
∑
( s , t )∈S XY
g r ( s, t )

● Useful for combination noise such as salt-and-pepper with

Gaussian noise
32
Example: Mean Filters

33
 Adaptive Filters
● Filters that consider behavior changes based on statistical
characteristics of the image inside the filter region.
● Capable of superior performance, but causes increase in filter
complexity
● Textbook Extra Reading:
– Adaptive, local noise reduction filter
– Adaptive median filter

34
 Adaptive Filters
● Filters that consider behavior changes based on statistical
characteristics of the image inside the filter region.
● Capable of superior performance, but causes increase in filter
complexity
● Textbook Extra Reading:
– Adaptive, local noise reduction filter
– Adaptive median filter

35
 Periodic Noise
● Periodic Noise can be effectively filtered using frequency
domain techniques
● Periodic Noise: Concentrated bursts of energy in the Fourier
transform, at locations corresponding to the frequencies of
the periodic interference
● Approach: Use selective filters to isolate noise – Bandreject,
Bandpass, Notch filters

36
 Bandreject Filters

● Bandreject Filters – Remove noise from a certain location (or band) in

the frequency domain
– Image corrupted with additive periodic noise can be easily removed
with a bandreject filter
● Bandpass Filter – the complement function of the Bandreject filter,
performs the opposite operation.
– Useful for isolating noise pattern for analysis
37
Example: Bandreject Filter

38
 Notch Filters
● Notch Filter – Rejects or passes frequencies in predefined
neighborhoods about a center frequency
● Due to symmetry of the Fourier Transform, notch filters must
appear in symmetric pairs about the origin in order to obtain
meaningful results
● Available as Notch Pass and Notch Reject (one a complement
of the other)
● Useful for removing periodic noise (horizontal, vertical,
diagonal periodic lines in images) which are concentrated on
one small spot in the frequency spectrum
39
Example: Notch (Reject) Filters

40
Example: Notch Filter

41
 Restoration Process Model
Degraded image
g(x,y)

G(u,v)
Frequency
Degradation function Fourier H(u,v) Domain
h(x, y) Transform Filter
R(u,v)
N(u,v)
Noise model
n(x, y)

Inverse
Restored Image
fˆ ( x, y ) Fourier
Transform

42
 Restoration Process
● Mathematical model:
G(u,v) = H(u,v)F(u,v) + N(u,v)
where:
G(u,v) = Fourier transform of degraded image
H(u,v) = Fourier transform of degradation function
F(u,v) = Fourier transform of original image
N(u,v) = Fourier transform of additive noise function
● To obtain the restored image:
fˆ ( x, y ) = ℑ−1 [ Fˆ (u, v)] = F −1 [ R(u, v)G (u, v)]
where:
fˆ ( x, y )
= the restored image, an approximation of
ℑ−1[] = the inverse Fourier transform
R(u,v) = the restoration (frequency domain) filter
43
 Degradation Function?
● Question: How do we estimate the degradation
function?
– Image Observation
– Experimentation
– Mathematical Modeling

44
 Inverse Filtering
● Uses the same model, with assumption of no noise.
Fourier transform of degraded image:
G (u , v) = H (u , v) F (u , v) + 0
● Fourier transform of the original image will be:
G (u , v) 1
F (u , v) = = G (u , v)
H (u, v) H (u , v)

● To find the original image, take the inverse Fourier

transform of F(u,v):
 G (u, v) 
f ( x, y ) = F −1 [ F (u, v)] = F −1  
 H (u, v) 
 1 
= F −1 G (u, v)
 H (u, v) 
45
 Example: Inverse Filtering

50 50 25  501 1

50
1
25 
=  201 
1
H (u , v) =  20 20 20 
1 1
H (u , v) 20 20 
 20 35 22   201 1
35
1
22


G (u , v) H (u , v) F (u , v) N (u , v) N (u , v)
Fˆ (u , v) = = + = F (u , v) +
H (u , v) H (u , v) H (u , v) H (u , v)

46
 Inverse Filtering: Some Problems
● If any points in H(u,v) are zero – division by zero
● Solution: Do not take zero-points of H(u,v) into account
● In presence of noise:
G (u, v) H (u , v) F (u, v) N (u, v) N (u, v)
Fˆ (u, v) = = + = F (u, v) +
H (u, v) H (u, v) H (u, v) H (u , v)
● As the value of H(u,v) becomes very small, the second term
becomes very large, and it overshadows the F(u,v)
● Limit the restoration to a specific radius about the origin in
the spectrum – the restoration cutoff frequency

47
Example: Inverse Filter
original image

Image blurred with

an 11 x 11 gaussian
convolution mask

Inverse filter, with cutoff

frequency = 40,
histogram stretched with 3%
low and high clipping to show
detail

Inverse filter, with cutoff

frequency = 60, histogram
stretched

48
 Wiener Filter
● Also known as minimum mean-square error (MMSE) filter
● Attempt to model the error in the restored image through
the use of statistical characteristics of noise
● The average error is mathematically minimized, resulting in
the equation for Wiener filter:
2
H * (u , v) 1 H (u , v)
Rw (u , v) = =
H (u , v) +  Snl ( u ,v )  H (u , v) H (u , v) +  Snl (u ,v ) 
2 S ( u ,v ) 2 S (u ,v )

where H * (u , v) = complex conjugate of H (u , v)

S n (u , v) = N (u , v) = power spectrum of the noise
2

Sl (u , v) = F (u , v) = power spectrum of the original image

2

49
Example: Wiener Filter
Image blurred with
an 11 x 11 gaussian
convolution mask

Image with gaussian noise

variance = 5; mean = 0

Inverse filter, with cutoff

frequency = 80, histogram
stretched with 3 % low and
high clipping to show detail

Wiener filter, with cutoff

frequency = 80, histogram
stretched

50
Example: Motion Blur + Additive Noise

51
 Other Frequency-Domain Restoration
Filters
● Constrained Least-Squares Filter
● Geometric Mean Filter – the most general form for frequency
domain restoration filters

52
 Recommended Readings
rd
● Digital Image Processing (3 Edition), Gonzalez &
Woods,
● Chapter 4: Image Restoration
● 5.1 – 5.4, 5.6 – 5.10 (Week 7)
● Chapter 9: Morphological Image Processing
● 9.1 – 9.4 (Week 8)

53