In the Name God

Digital Image Processing

Chapter 5 - Image Restoration

and Reconstruction
[email protected] Abbas Koochari
Preview
Restoration is an objective process compared to image enhancement:
Image restoration is to restore a degraded image back to the original image.
Image Enhancement is to manipulate the image so that it is suitable for a
specific application.
Contrast stretching is an enhancement technique while debluring function is
considered a restoration.
Only consider in this chapter a degraded digital image.
Restoration can be categorized as two groups:
Deterministic methods are applicable to images with little noise and a known
degradation
Stochastic methods try to find the best restoration according to a particular
stochastic criterion, e.g., a least square method

2
 5.1 A Model of the Image
Degradation/Restoration Process

3
A Model of the Image
Degradation/Restoration Process

5.1- A Model of the Image

Degradation/Restoration Process

5.2- Noise Models

5.3- Restoration in the Presence
of Noise Only-Spatial Filtering
5.4- Periodic Noise Reduction by
Frequency Domain Filtering
5.5 - Linear, Position-Invariant
Degradations
5.6- Estimating the degradation
Function
5.7- Inverse Filtering
5.8- Minimum Mean Square
Error (Wiener) Filtering
5.9- Constrained Least Square
Filtering
5.10- Geometric Mean Filter
5.11- Image Reconstruction from
Projections

4
A Model of the Image
Degradation/Restoration Process

5.1- A Model of the Image

Degradation/Restoration Process Spatial domain: additive noise
The degraded image in Spatial domain is
5.2- Noise Models
5.3- Restoration in the Presence g ( x, y )  f ( x, y )  h ( x, y )   ( x , y )
of Noise Only-Spatial Filtering
5.4- Periodic Noise Reduction by
Frequency Domain Filtering
where h(x,y) is a system that causes image
5.5 - Linear, Position-Invariant distortion and ƞ(x,y) is noise.
Degradations
5.6- Estimating the degradation
Function Frequency domain : blurring
5.7- Inverse Filtering The degraded image in Frequency domain is
5.8- Minimum Mean Square
Error (Wiener) Filtering G (u , v)  H (u , v) F (u , v)  N (u , v)
5.9- Constrained Least Square
Filtering
Where the terms in capital letters are Fourier
5.10- Geometric Mean Filter
5.11- Image Reconstruction from
transforms.
Projections

5
Noise Models

5.1- A Model of the Image

Degradation/Restoration Process

5.2- Noise Models

Spatial and Frequency Properties of Noise
5.3- Restoration in the Presence
of Noise Only-Spatial Filtering
5.4- Periodic Noise Reduction by
Frequency Domain Filtering Some Important Noise Probability Density Functions
5.5 - Linear, Position-Invariant
Degradations
5.6- Estimating the degradation
Function Periodic Noise
5.7- Inverse Filtering
5.8- Minimum Mean Square
Error (Wiener) Filtering
5.9- Constrained Least Square
Estimation of Noise Parameters
Filtering
5.10- Geometric Mean Filter
5.11- Image Reconstruction from
Projections
6
The Principal Source of Noise

5.1- A Model of the Image

Degradation/Restoration Process

5.2- Noise Models

Noise arise …
5.3- Restoration in the Presence
of Noise Only-Spatial Filtering During Image Acquisition
5.4- Periodic Noise Reduction by Environment conditions
Frequency Domain Filtering
5.5 - Linear, Position-Invariant
Quality of sensing elements
Degradations For x. Two factors for CCD: light level and
5.6- Estimating the degradation sensor temperature
Function
5.7- Inverse Filtering
5.8- Minimum Mean Square Image Transmission
Error (Wiener) Filtering
5.9- Constrained Least Square
Filtering
5.10- Geometric Mean Filter
5.11- Image Reconstruction from
Projections
7
Spatial and Frequency Properties of
Noise

5.1- A Model of the Image

Degradation/Restoration Process

5.2- Noise Models

White noise:
5.3- Restoration in the Presence The Fourier spectrum of noise is constant.
of Noise Only-Spatial Filtering This terminology is a carryover from the physical
5.4- Periodic Noise Reduction by properties of white light, which contains nearly all
Frequency Domain Filtering
5.5 - Linear, Position-Invariant
frequencies in the visible spectrum in equal properties.
Degradations
5.6- Estimating the degradation
Function We assume in this chapter: Noise is independent of spatial
5.7- Inverse Filtering coordinates.
5.8- Minimum Mean Square
Error (Wiener) Filtering
5.9- Constrained Least Square
Filtering
5.10- Geometric Mean Filter
5.11- Image Reconstruction from
Projections
8
Noise Probability Density Functions

5.1- A Model of the Image

Degradation/Restoration Process

5.2- Noise Models

Noise cannot be predicted but can be approximately
5.3- Restoration in the Presence
of Noise Only-Spatial Filtering described in statistical way using the probability density
5.4- Periodic Noise Reduction by function (PDF).
Frequency Domain Filtering
5.5 - Linear, Position-Invariant
Degradations The statistical properties of the gray level of spatial noise
5.6- Estimating the degradation can be considered random variables characterized by a PDF.
Function
5.7- Inverse Filtering
5.8- Minimum Mean Square
Error (Wiener) Filtering
5.9- Constrained Least Square
Filtering
5.10- Geometric Mean Filter
5.11- Image Reconstruction from
Projections
9
Most Common PDFs of Noises

5.1- A Model of the Image

Degradation/Restoration Process Gaussian noise
Are used frequently in practice
5.2- Noise Models The PDF of a Gaussian random variable, Z, is given by:
5.3- Restoration in the Presence
1
of Noise Only-Spatial Filtering
p( z )  e ( z   ) / 2 2
2

5.4- Periodic Noise Reduction by

Rayleigh noise 2
Frequency Domain Filtering
5.5 - Linear, Position-Invariant The PDF of Rayleigh noise:
Degradations
5.6- Estimating the degradation
Function  2
 ( z  a )e ( z  a ) / b for z  a
2

5.7- Inverse Filtering p( z )   b

5.8- Minimum Mean Square 
0 for z  a
Error (Wiener) Filtering Erlang (Gamma) noise
5.9- Constrained Least Square The PDF of Erlang noise :
Filtering
 b b 1 ( z  a ) 2 / b
a z
5.10- Geometric Mean Filter
5.11- Image Reconstruction from p ( z )   (b  1)! e for z  a
Projections
10 0 for z  a

Most Common PDFs of Noises

5.1- A Model of the Image Exponential noise

Degradation/Restoration Process The PDF of exponential noise : p ( z )  ae  az
5.2- Noise Models
Uniform noise  1
5.3- Restoration in the Presence The PDF of uniform noise is given by: p ( z )  
b - a
for a  z  b
of Noise Only-Spatial Filtering 
 0 otherwise
5.4- Periodic Noise Reduction by
Frequency Domain Filtering
5.5 - Linear, Position-Invariant Impulse noise (Salt and pepper)  Pa for z  a
Degradations The PDF of impulse noise is given by: p( z )  
 Pb for z  b
5.6- Estimating the degradation 0
Function  otherwise
5.7- Inverse Filtering
5.8- Minimum Mean Square
Error (Wiener) Filtering If b>a gray level b will appear as a light dot;
5.9- Constrained Least Square If either Pa or Pb is zero, the impulse is called unipolar
Filtering
5.10- Geometric Mean Filter
If neither probability is zero (bipolar), and especially if they are
approximately equal: salt and pepper noise
5.11- Image Reconstruction from
Projections
11
Most Common PDFs of Noises

5.1- A Model of the Image

Degradation/Restoration Process

5.2- Noise Models

5.3- Restoration in the Presence

of Noise Only-Spatial Filtering
5.4- Periodic Noise Reduction by
Frequency Domain Filtering
5.5 - Linear, Position-Invariant
Degradations
5.6- Estimating the degradation PDF tells
Function how much
5.7- Inverse Filtering each z value
5.8- Minimum Mean Square occurs.
Error (Wiener) Filtering
5.9- Constrained Least Square
Filtering
5.10- Geometric Mean Filter
5.11- Image Reconstruction from
Projections
12
Noise Factors

5.1- A Model of the Image

Degradation/Restoration Process

5.2- Noise Models Gaussian noise: electronic circuit noise and sensors noise due
to poor illumination and /or temperature
5.3- Restoration in the Presence
of Noise Only-Spatial Filtering Rayleigh noise: helpful in characterizing noise phenomena in
5.4- Periodic Noise Reduction by ranging image
Frequency Domain Filtering Exponential and gamma noise: application in laser imaging
5.5 - Linear, Position-Invariant
Degradations Impulse noise: found in quick transient such as faulty-
5.6- Estimating the degradation switching ; is the only one that is visually indicative
Function
Uniform noise: basis for random number generator
5.7- Inverse Filtering
5.8- Minimum Mean Square Difficult to differentiate visually between the five image
Error (Wiener) Filtering (Fig 5.4(a) ~Fig5.4(b))
5.9- Constrained Least Square
Filtering
5.10- Geometric Mean Filter
5.11- Image Reconstruction from
Projections
13
Image Degradation with Additive Noise

5.1- A Model of the Image

Degradation/Restoration Process g ( x, y )  f ( x, y )   ( x, y )
5.2- Noise Models Original image Degraded images

5.3- Restoration in the Presence

of Noise Only-Spatial Filtering
5.4- Periodic Noise Reduction by
Frequency Domain Filtering
5.5 - Linear, Position-Invariant
Degradations
5.6- Estimating the degradation
Function
5.7- Inverse Filtering
5.8- Minimum Mean Square
Error (Wiener) Filtering Histogram
5.9- Constrained Least Square
Filtering
5.10- Geometric Mean Filter
5.11- Image Reconstruction from
Projections
14
Image Degradation with Additive Noise

5.1- A Model of the Image

Degradation/Restoration Process g ( x, y )  f ( x, y )   ( x, y )

5.2- Noise Models

Original image Degraded images
5.3- Restoration in the Presence
of Noise Only-Spatial Filtering
5.4- Periodic Noise Reduction by
Frequency Domain Filtering
5.5 - Linear, Position-Invariant
Degradations
5.6- Estimating the degradation
Function
5.7- Inverse Filtering
5.8- Minimum Mean Square
Error (Wiener) Filtering
5.9- Constrained Least Square
Filtering Histogram
5.10- Geometric Mean Filter
5.11- Image Reconstruction from
Projections
15
Image Degradation with Additive Noise

5.1- A Model of the Image

Degradation/Restoration Process

5.2- Noise Models

5.3- Restoration in the Presence

of Noise Only-Spatial Filtering
5.4- Periodic Noise Reduction by
Frequency Domain Filtering
5.5 - Linear, Position-Invariant
Degradations
5.6- Estimating the degradation
Function
5.7- Inverse Filtering
5.8- Minimum Mean Square
Error (Wiener) Filtering
5.9- Constrained Least Square
Filtering
5.10- Geometric Mean Filter
5.11- Image Reconstruction from
Projections
16
Image Degradation with Additive Noise

5.1- A Model of the Image

Degradation/Restoration Process

5.2- Noise Models

5.3- Restoration in the Presence

of Noise Only-Spatial Filtering
5.4- Periodic Noise Reduction by
Frequency Domain Filtering
5.5 - Linear, Position-Invariant
Degradations
5.6- Estimating the degradation
Function
5.7- Inverse Filtering
5.8- Minimum Mean Square
Error (Wiener) Filtering
5.9- Constrained Least Square
Filtering
5.10- Geometric Mean Filter
5.11- Image Reconstruction from
Projections
17
Image Degradation with Additive Noise

5.1- A Model of the Image

Degradation/Restoration Process

5.2- Noise Models

5.3- Restoration in the Presence

of Noise Only-Spatial Filtering
5.4- Periodic Noise Reduction by
Frequency Domain Filtering
5.5 - Linear, Position-Invariant
Degradations
5.6- Estimating the degradation
Function
5.7- Inverse Filtering
5.8- Minimum Mean Square
Error (Wiener) Filtering
5.9- Constrained Least Square
Filtering
5.10- Geometric Mean Filter
5.11- Image Reconstruction from
Projections
18
Periodic Noise

5.1- A Model of the Image

Degradation/Restoration Process

5.2- Noise Models

Caused by electrical or electromechanical interference during
5.3- Restoration in the Presence image acquisition
of Noise Only-Spatial Filtering
5.4- Periodic Noise Reduction by
Frequency Domain Filtering Can be reduced significantly via frequency domain filtering
5.5 - Linear, Position-Invariant (For example: Sinusoidal wave, the impulses appear in an
Degradations
5.6- Estimating the degradation
approximate circle)
Function
5.7- Inverse Filtering
For Ex. (Fig. 5.5): An image severely corrupted by sinusoidal
5.8- Minimum Mean Square
Error (Wiener) Filtering noise of various frequency
5.9- Constrained Least Square
Filtering
5.10- Geometric Mean Filter
5.11- Image Reconstruction from
Projections
19
Periodic Noise

5.1- A Model of the Image

Degradation/Restoration Process

5.2- Noise Models

5.3- Restoration in the Presence

of Noise Only-Spatial Filtering
5.4- Periodic Noise Reduction by
Frequency Domain Filtering
5.5 - Linear, Position-Invariant
Degradations Noise components
(Periodic noise
5.6- Estimating the degradation
Function looks like dots In
the frequency
5.7- Inverse Filtering
Domain)
5.8- Minimum Mean Square
Error (Wiener) Filtering
5.9- Constrained Least Square Periodic noise can
Filtering be reduced in via
5.10- Geometric Mean Filter frequency domain
5.11- Image Reconstruction from
Projections Are generated due to electrical or electromechanical interference during
20 image acquisition
Estimation of Noise Parameters

5.1- A Model of the Image

Degradation/Restoration Process

5.2- Noise Models The parameters of periodic noise are often estimated by
inspection of the Fourier spectrum of the image.
5.3- Restoration in the Presence Periodicity tends to produce frequency spikes that often can
of Noise Only-Spatial Filtering
5.4- Periodic Noise Reduction by be detected even by visual analysis
Frequency Domain Filtering Attempt to infer the periodicity of noise components
5.5 - Linear, Position-Invariant
Degradations
5.6- Estimating the degradation How to identify the noise model?
Function
5.7- Inverse Filtering
Calculate the mean and variance of the gray levels
5.8- Minimum Mean Square
(illustrate with the image strips in Fig. 5.4)
Error (Wiener) Filtering The shape of the histogram identifies the closet PDF
5.9- Constrained Least Square match
Filtering
5.10- Geometric Mean Filter
5.11- Image Reconstruction from
Projections
21
Estimation of Noise Parameters

5.1- A Model of the Image

Degradation/Restoration Process

5.2- Noise Models

5.3- Restoration in the Presence

of Noise Only-Spatial Filtering We cannot use the
5.4- Periodic Noise Reduction by image histogram to
Frequency Domain Filtering
5.5 - Linear, Position-Invariant estimate noise PDF.
Degradations
5.6- Estimating the degradation
Function
5.7- Inverse Filtering
It is better to use the
5.8- Minimum Mean Square histogram of one area
Error (Wiener) Filtering of an image that has
5.9- Constrained Least Square constant intensity to
Filtering
5.10- Geometric Mean Filter estimate noise PDF.
5.11- Image Reconstruction from
Projections
22
5.3 Restoration in the Presence of
Noise Only - Spatial Filtering

23
Restoration in the Presence of Noise
Only - Spatial Filtering

5.1- A Model of the Image

Degradation/Restoration Process
5.2- Noise Models

5.3- Restoration in the Presence Mean Filters

of Noise Only-Spatial Filtering

5.4- Periodic Noise Reduction by

Frequency Domain Filtering Ordered-Statistic Filters
5.5 - Linear, Position-Invariant
Degradations
5.6- Estimating the degradation
Function Adaptive Filters
5.7- Inverse Filtering
5.8- Minimum Mean Square
Error (Wiener) Filtering
5.9- Constrained Least Square
Filtering
5.10- Geometric Mean Filter
5.11- Image Reconstruction from
Projections
24
Mean Filters

5.1- A Model of the Image

Degradation/Restoration Process
5.2- Noise Models Performance superior to the filters discussed in Section 3.6 Degradation Model:

5.3- Restoration in the Presence

of Noise Only-Spatial Filtering g ( x, y )  f ( x, y )  h( x, y )   ( x, y )
5.4- Periodic Noise Reduction by To remove this part
Frequency Domain Filtering
5.5 - Linear, Position-Invariant Arithmetic Mean Filter(Moving Average Filter):
Degradations Computes the average value of the corrupted image g(x,y)
5.6- Estimating the degradation The value of the restored image f
Function
5.7- Inverse Filtering
5.8- Minimum Mean Square
1
Error (Wiener) Filtering
5.9- Constrained Least Square
fˆ ( x, y )   g ( s, t )
mn ( s ,t )S xy
Filtering
5.10- Geometric Mean Filter mn = size of moving window
5.11- Image Reconstruction from
Projections
25
Mean Filters

5.1- A Model of the Image

Degradation/Restoration Process Geometric Mean Filter:
5.2- Noise Models Smooth comparable to the arithmetic mean filter,
but it tends to loss less detail.
5.3- Restoration in the Presence
of Noise Only-Spatial Filtering 1
  mn
fˆ ( x, y )    g ( s, t ) 
5.4- Periodic Noise Reduction by  ( s ,t )S 
Frequency Domain Filtering  xy 
5.5 - Linear, Position-Invariant
Degradations Harmonic Mean Filter:
5.6- Estimating the degradation
Function Work well for salt noise and Gaussian noise, but
5.7- Inverse Filtering fails for pepper noise
5.8- Minimum Mean Square
Error (Wiener) Filtering mn
5.9- Constrained Least Square fˆ ( x, y ) 
1
Filtering 
( s ,t )S xy g ( s, t )
Works well for salt noise
but fails for pepper noise
5.10- Geometric Mean Filter
5.11- Image Reconstruction from
Projections
26
Mean Filters

5.1- A Model of the Image

Degradation/Restoration Process
5.2- Noise Models Contra harmonic Mean Filter:

5.3- Restoration in the Presence

of Noise Only-Spatial Filtering

5.4- Periodic Noise Reduction by  g ( s, t ) Q 1

Positive Q is suitable for
fˆ ( x, y ) 
Frequency Domain Filtering ( s ,t )S xy
eliminating pepper noise.
 g ( s, t )
5.5 - Linear, Position-Invariant Q
Degradations Negative Q is suitable for
5.6- Estimating the degradation ( s ,t )S xy eliminating salt noise.
Function
5.7- Inverse Filtering Q = the filter order
5.8- Minimum Mean Square
Error (Wiener) Filtering
5.9- Constrained Least Square For Q = 0, the filter reduces to an arithmetic mean filter.
Filtering For Q = -1, the filter reduces to a harmonic mean filter.
5.10- Geometric Mean Filter
5.11- Image Reconstruction from
Projections
27
Arithmetic Mean Filter - Example

5.1- A Model of the Image

Degradation/Restoration Process
5.2- Noise Models

5.3- Restoration in the Presence

of Noise Only-Spatial Filtering

5.4- Periodic Noise Reduction by

Frequency Domain Filtering
5.5 - Linear, Position-Invariant
Degradations
5.6- Estimating the degradation
Function
5.7- Inverse Filtering
5.8- Minimum Mean Square
Error (Wiener) Filtering
5.9- Constrained Least Square
Filtering
5.10- Geometric Mean Filter
5.11- Image Reconstruction from
Projections
28
Arithmetic Mean Filter - Example

5.1- A Model of the Image

Degradation/Restoration Process
5.2- Noise Models

5.3- Restoration in the Presence

of Noise Only-Spatial Filtering

5.4- Periodic Noise Reduction by

Frequency Domain Filtering
5.5 - Linear, Position-Invariant
Degradations
5.6- Estimating the degradation
Function
5.7- Inverse Filtering
5.8- Minimum Mean Square
Error (Wiener) Filtering
5.9- Constrained Least Square
Filtering
5.10- Geometric Mean Filter
5.11- Image Reconstruction from
Projections
29
Arithmetic Mean Filter - Example

5.1- A Model of the Image

Degradation/Restoration Process
5.2- Noise Models

5.3- Restoration in the Presence

of Noise Only-Spatial Filtering

5.4- Periodic Noise Reduction by

Frequency Domain Filtering
5.5 - Linear, Position-Invariant
Degradations
5.6- Estimating the degradation
Function
5.7- Inverse Filtering
5.8- Minimum Mean Square
Error (Wiener) Filtering
5.9- Constrained Least Square
Filtering
5.10- Geometric Mean Filter
5.11- Image Reconstruction from
Projections
30
Geometric Mean Filter - Example

5.1- A Model of the Image

Degradation/Restoration Process
5.2- Noise Models

5.3- Restoration in the Presence Image

Original
of Noise Only-Spatial Filtering Corrupted
Image
by AWGN
5.4- Periodic Noise Reduction by
Frequency Domain Filtering
5.5 - Linear, Position-Invariant
Degradations
5.6- Estimating the degradation Image
Function obtained Image
5.7- Inverse Filtering using a 3x3 obtained
5.8- Minimum Mean Square arithmetic using a 3x3
Error (Wiener) Filtering mean filter geometric
5.9- Constrained Least Square mean filter
Filtering
5.10- Geometric Mean Filter
5.11- Image Reconstruction from AWGN: Additive White Gaussian Noise
Projections
31
Geometric Mean Filter - Example

5.1- A Model of the Image

Degradation/Restoration Process
5.2- Noise Models

5.3- Restoration in the Presence

of Noise Only-Spatial Filtering Image corrupted Image corrupted
by pepper noise by salt noise with
5.4- Periodic Noise Reduction by with prob. = 0.1
Frequency Domain Filtering prob. = 0.1
5.5 - Linear, Position-Invariant
Degradations
5.6- Estimating the degradation Image obtained
Function using a 3x3
5.7- Inverse Filtering Image obtained
contra-
5.8- Minimum Mean Square using a 3x3
harmonic mean
Error (Wiener) Filtering contra-harmonic
filter
5.9- Constrained Least Square mean filter With
With Q = 1.5
Filtering Q=-1.5
5.10- Geometric Mean Filter
5.11- Image Reconstruction from
Projections
32
Contraharmonic Filters: Incorrect Use

5.1- A Model of the Image

Degradation/Restoration Process
5.2- Noise Models

5.3- Restoration in the Presence Image

corrupted Image
of Noise Only-Spatial Filtering corrupted
by pepper
noise with by salt
5.4- Periodic Noise Reduction by prob. = 0.1 noise with
Frequency Domain Filtering prob. = 0.1
5.5 - Linear, Position-Invariant
Degradations
5.6- Estimating the degradation
Image
Function obtained
5.7- Inverse Filtering using a 3x3 Image
contra- obtained
5.8- Minimum Mean Square using a 3x3
Error (Wiener) Filtering harmonic
mean filter contra-
5.9- Constrained Least Square With Q=-1.5 harmonic
Filtering mean filter
5.10- Geometric Mean Filter With Q=1.5

5.11- Image Reconstruction from

Projections
33
Contra harmonic Filters - Example

5.1- A Model of the Image

Degradation/Restoration Process
The Contraharmonic mean filter works well for images
5.2- Noise Models
containing salt or pepper type noise
5.3- Restoration in the Presence
of Noise Only-Spatial Filtering

5.4- Periodic Noise Reduction by

Frequency Domain Filtering
5.5 - Linear, Position-Invariant
Degradations
5.6- Estimating the degradation
Function
5.7- Inverse Filtering
5.8- Minimum Mean Square
Error (Wiener) Filtering
5.9- Constrained Least Square
Filtering
5.10- Geometric Mean Filter
5.11- Image Reconstruction from
Projections
34
Contra harmonic Filters - Example

5.1- A Model of the Image

Degradation/Restoration Process
5.2- Noise Models

5.3- Restoration in the Presence

of Noise Only-Spatial Filtering

5.4- Periodic Noise Reduction by

Frequency Domain Filtering
5.5 - Linear, Position-Invariant
Degradations
5.6- Estimating the degradation
Function
5.7- Inverse Filtering
5.8- Minimum Mean Square
Error (Wiener) Filtering
5.9- Constrained Least Square
Filtering
5.10- Geometric Mean Filter
5.11- Image Reconstruction from
Projections
35
Order-Statistic Filters: Revisit

5.1- A Model of the Image Subimage

Degradation/Restoration Process
5.2- Noise Models
Original image
5.3- Restoration in the Presence
of Noise Only-Spatial Filtering

5.4- Periodic Noise Reduction by Statistic parameters

Frequency Domain Filtering Mean, Median, Mode,
5.5 - Linear, Position-Invariant Min, Max, Etc.
Degradations
5.6- Estimating the degradation
Function
5.7- Inverse Filtering
5.8- Minimum Mean Square
Error (Wiener) Filtering
5.9- Constrained Least Square Moving
Filtering
Window
5.10- Geometric Mean Filter
5.11- Image Reconstruction from
Projections Output image
36
Order-Statistics Filters

5.1- A Model of the Image

Median filters:
Degradation/Restoration Process Are particularly effective in the presence of both bipolar
5.2- Noise Models and unipolar impulse noise
5.3- Restoration in the Presence fˆ ( x, y )  median g ( s, t )
of Noise Only-Spatial Filtering ( s ,t )S xy

5.4- Periodic Noise Reduction by

Frequency Domain Filtering Max filters (Fig. 5.8)
5.5 - Linear, Position-Invariant Max filter: reduce low values caused by pepper noise
Degradations
5.6- Estimating the degradation
Function fˆ ( x, y )  max g ( s, t ) Reduce “dark” noise
5.7- Inverse Filtering ( s ,t )S xy (pepper noise)
5.8- Minimum Mean Square
Error (Wiener) Filtering Min filters (Fig. 5.8)
5.9- Constrained Least Square
Filtering
Min filter: reduce high values caused by salt noise
5.10- Geometric Mean Filter
5.11- Image Reconstruction from fˆ ( x, y )  min g ( s, t ) Reduce “bright”
Projections ( s ,t )S xy
37 noise (salt noise)
Order-Statistics Filters

5.1- A Model of the Image

Degradation/Restoration Process
5.2- Noise Models
Midpoint filter
5.3- Restoration in the Presence
of Noise Only-Spatial Filtering
Combines order statistics and averaging

5.4- Periodic Noise Reduction by

Frequency Domain Filtering
5.5 - Linear, Position-Invariant
1 
fˆ ( x, y )   max g ( s, t ) min g ( s, t )
Degradations
5.6- Estimating the degradation
Function 2  ( s ,t )S xy ( s ,t )S xy 
5.7- Inverse Filtering
5.8- Minimum Mean Square
Error (Wiener) Filtering
5.9- Constrained Least Square
Filtering
5.10- Geometric Mean Filter
5.11- Image Reconstruction from
Projections
38
Median Filter: How it works

A median filter is good for removing impulse, isolated noise

5.1- A Model of the Image
Degradation/Restoration Process
5.2- Noise Models Salt noise
Pepper noise
5.3- Restoration in the Presence Median
of Noise Only-Spatial Filtering

Moving
5.4- Periodic Noise Reduction by Degraded Window
Frequency Domain Filtering image
5.5 - Linear, Position-Invariant Sorted
Degradations Salt noise Array
5.6- Estimating the degradation
Function Pepper noise
5.7- Inverse Filtering
Filter output
5.8- Minimum Mean Square
Error (Wiener) Filtering
Normally, impulse noise has high magnitude and is
5.9- Constrained Least Square
Filtering solated. When we sort pixels in the moving window,
5.10- Geometric Mean Filter noise pixels are usually at the ends of the array.
5.11- Image Reconstruction from
Projections Therefore, it’s rare that the noise pixel will be a median value.
39
Median Filter: How it works

5.1- A Model of the Image

Degradation/Restoration Process
5.2- Noise Models
1 2
5.3- Restoration in the Presence
of Noise Only-Spatial Filtering Image
corrupted by
5.4- Periodic Noise Reduction by
Frequency Domain Filtering
salt and pepper
5.5 - Linear, Position-Invariant noise with
Degradations pa=pb= 0.1
5.6- Estimating the degradation
Function
3 4
5.7- Inverse Filtering
5.8- Minimum Mean Square
Error (Wiener) Filtering
5.9- Constrained Least Square
Filtering
5.10- Geometric Mean Filter
5.11- Image Reconstruction from
Projections Images obtained using a 3x3 median filter
40
Median Filter: How it works

5.1- A Model of the Image

Degradation/Restoration Process
5.2- Noise Models

5.3- Restoration in the Presence

of Noise Only-Spatial Filtering

5.4- Periodic Noise Reduction by

Frequency Domain Filtering
5.5 - Linear, Position-Invariant
Degradations
5.6- Estimating the degradation
Function
5.7- Inverse Filtering
5.8- Minimum Mean Square
Error (Wiener) Filtering
5.9- Constrained Least Square
Filtering
5.10- Geometric Mean Filter
5.11- Image Reconstruction from
Projections
41
Median Filter: How it works

5.1- A Model of the Image

Degradation/Restoration Process
5.2- Noise Models

5.3- Restoration in the Presence

of Noise Only-Spatial Filtering

5.4- Periodic Noise Reduction by

Frequency Domain Filtering
5.5 - Linear, Position-Invariant
Degradations
5.6- Estimating the degradation
Function
5.7- Inverse Filtering
5.8- Minimum Mean Square
Error (Wiener) Filtering
5.9- Constrained Least Square
Filtering
5.10- Geometric Mean Filter
5.11- Image Reconstruction from
Projections
42
Max and Min Filters - Example

5.1- A Model of the Image

Degradation/Restoration Process
5.2- Noise Models

5.3- Restoration in the Presence Image

of Noise Only-Spatial Filtering Image
corrupted corrupted
5.4- Periodic Noise Reduction by by pepper by salt
Frequency Domain Filtering noise with noise with
5.5 - Linear, Position-Invariant prob. = 0.1 prob. = 0.1
Degradations
5.6- Estimating the degradation
Function Image
5.7- Inverse Filtering Image obtained
5.8- Minimum Mean Square obtained using a
Error (Wiener) Filtering using a 3x3 3x3
5.9- Constrained Least Square max filter min filter
Filtering
5.10- Geometric Mean Filter
5.11- Image Reconstruction from
Projections
43
Max and Min Filters - Example

5.1- A Model of the Image

Degradation/Restoration Process
5.2- Noise Models Various Windows size for Max and Min Filters:
5.3- Restoration in the Presence
of Noise Only-Spatial Filtering

5.4- Periodic Noise Reduction by

Frequency Domain Filtering
5.5 - Linear, Position-Invariant
Degradations
5.6- Estimating the degradation
Function
5.7- Inverse Filtering
5.8- Minimum Mean Square
Error (Wiener) Filtering
5.9- Constrained Least Square
Filtering
5.10- Geometric Mean Filter
5.11- Image Reconstruction from
Projections
44
Max and Min Filters - Example

5.1- A Model of the Image

Degradation/Restoration Process
5.2- Noise Models

5.3- Restoration in the Presence

of Noise Only-Spatial Filtering

5.4- Periodic Noise Reduction by

Frequency Domain Filtering
5.5 - Linear, Position-Invariant
Degradations
5.6- Estimating the degradation
Function
5.7- Inverse Filtering
5.8- Minimum Mean Square
Error (Wiener) Filtering
5.9- Constrained Least Square
Filtering
5.10- Geometric Mean Filter
5.11- Image Reconstruction from
Projections
45
Max and Min Filters - Example

5.1- A Model of the Image

Degradation/Restoration Process
5.2- Noise Models

5.3- Restoration in the Presence

of Noise Only-Spatial Filtering

5.4- Periodic Noise Reduction by

Frequency Domain Filtering
5.5 - Linear, Position-Invariant
Degradations
5.6- Estimating the degradation
Function
5.7- Inverse Filtering
5.8- Minimum Mean Square
Error (Wiener) Filtering
5.9- Constrained Least Square
Filtering
5.10- Geometric Mean Filter
5.11- Image Reconstruction from
Projections
46
Max and Min Filters - Example

5.1- A Model of the Image

Degradation/Restoration Process
5.2- Noise Models

5.3- Restoration in the Presence

of Noise Only-Spatial Filtering

5.4- Periodic Noise Reduction by

Frequency Domain Filtering
5.5 - Linear, Position-Invariant
Degradations
5.6- Estimating the degradation
Function
5.7- Inverse Filtering
5.8- Minimum Mean Square
Error (Wiener) Filtering
5.9- Constrained Least Square
Filtering
5.10- Geometric Mean Filter
5.11- Image Reconstruction from
Projections
47
Alpha -trimmed Mean Filter

5.1- A Model of the Image

Degradation/Restoration Process Formula:
5.2- Noise Models
1
5.3- Restoration in the Presence
of Noise Only-Spatial Filtering
fˆ ( x, y ) 
mn  d
 g ( s, t )
( s ,t )S xy
r

5.4- Periodic Noise Reduction by

Frequency Domain Filtering
5.5 - Linear, Position-Invariant
Degradations where gr(s,t) represent the remaining mn-d pixels
5.6- Estimating the degradation after removing the d/2 highest and d/2 lowest
Function values of g(s,t).
5.7- Inverse Filtering
5.8- Minimum Mean Square
Error (Wiener) Filtering This filter is useful in situations involving multiple types
5.9- Constrained Least Square of noise such as a combination of salt-and-pepper and
Filtering Gaussian noise.
5.10- Geometric Mean Filter
5.11- Image Reconstruction from
Projections
48
Alpha-trimmed Mean Filter - Example

5.1- A Model of the Image

Degradation/Restoration Process
5.2- Noise Models

5.3- Restoration in the Presence Image Image

of Noise Only-Spatial Filtering corrupted 1 2 additionally
by additive corrupted
5.4- Periodic Noise Reduction by by additive
Frequency Domain Filtering uniform
5.5 - Linear, Position-Invariant noise salt-and-
Degradations pepper
5.6- Estimating the degradation noise
Function Image 2
5.7- Inverse Filtering
obtained Image 2
5.8- Minimum Mean Square
using a 5x5 obtained
Error (Wiener) Filtering using a 5x5
5.9- Constrained Least Square arithmetic
mean filter geometric
Filtering
5.10- Geometric Mean Filter
mean filter
5.11- Image Reconstruction from
Projections
49
Alpha-trimmed Mean Filter - Example

5.1- A Model of the Image

Degradation/Restoration Process
5.2- Noise Models
Image
5.3- Restoration in the Presence Image 1 2 additionally
of Noise Only-Spatial Filtering corrupted corrupted
by additive by additive
5.4- Periodic Noise Reduction by uniform salt-and-
Frequency Domain Filtering
noise pepper
5.5 - Linear, Position-Invariant
Degradations noise
5.6- Estimating the degradation
Function Image 2
5.7- Inverse Filtering Image 2 obtained
5.8- Minimum Mean Square obtained using a 5x5
Error (Wiener) Filtering using a 5x5 alpha-
5.9- Constrained Least Square median filter trimmed
Filtering
5.10- Geometric Mean Filter
mean filter
5.11- Image Reconstruction from
with d = 5
Projections
50
Alpha-trimmed Mean Filter - Example

5.1- A Model of the Image

Degradation/Restoration Process
5.2- Noise Models
Image
5.3- Restoration in the Presence obtained Image
of Noise Only-Spatial Filtering using a 5x5 obtained
arithmetic using a 5x5
5.4- Periodic Noise Reduction by mean filter geometric
Frequency Domain Filtering mean filter
5.5 - Linear, Position-Invariant
Degradations
5.6- Estimating the degradation
Function Image
Image obtained
5.7- Inverse Filtering
obtained using a 5x5
5.8- Minimum Mean Square
using a 5x5 alpha-
Error (Wiener) Filtering
median filter trimmed
5.9- Constrained Least Square
mean filter
Filtering
with d = 5
5.10- Geometric Mean Filter
5.11- Image Reconstruction from
Projections
51
Alpha-trimmed Mean Filter - Example

5.1- A Model of the Image

Degradation/Restoration Process
5.2- Noise Models

5.3- Restoration in the Presence

of Noise Only-Spatial Filtering

5.4- Periodic Noise Reduction by

Frequency Domain Filtering
5.5 - Linear, Position-Invariant
Degradations
5.6- Estimating the degradation
Function
5.7- Inverse Filtering
5.8- Minimum Mean Square
Error (Wiener) Filtering
5.9- Constrained Least Square
Filtering
5.10- Geometric Mean Filter
5.11- Image Reconstruction from
Projections
52
Alpha-trimmed Mean Filter - Example

5.1- A Model of the Image

Degradation/Restoration Process
5.2- Noise Models

5.3- Restoration in the Presence

of Noise Only-Spatial Filtering

5.4- Periodic Noise Reduction by

Frequency Domain Filtering
5.5 - Linear, Position-Invariant
Degradations
5.6- Estimating the degradation
Function
5.7- Inverse Filtering
5.8- Minimum Mean Square
Error (Wiener) Filtering
5.9- Constrained Least Square
Filtering
5.10- Geometric Mean Filter
5.11- Image Reconstruction from
Projections
53
Adaptive Filters

5.1- A Model of the Image

Degradation/Restoration Process Global Concept:
5.2- Noise Models
Can apply to an image with regard to how image
5.3- Restoration in the Presence
characteristics vary from one point to another
of Noise Only-Spatial Filtering Filter behavior depends on statistical characteristics
of local areas inside m×n moving window
5.4- Periodic Noise Reduction by
Frequency Domain Filtering More complex but superior performance compared
5.5 - Linear, Position-Invariant with “fixed” filters
Degradations Statistical characteristics:
5.6- Estimating the degradation
Function
1  2
5.7- Inverse Filtering
5.8- Minimum Mean Square
Local mean: mL   g ( s, t )
mn ( s ,t )S xy
Noise variance
Error (Wiener) Filtering
5.9- Constrained Least Square
1

Filtering
5.10- Geometric Mean Filter
Local Variance:  L2  ( g ( s, t )  mL )2
mn ( s ,t )S xy
5.11- Image Reconstruction from
Projections
54
Adaptive, Local Noise Reduction
Filters

5.1- A Model of the Image The response of the filter based on four quantities:
Degradation/Restoration Process
The value of the noise image
5.2- Noise Models
The variance of the noise corrupting f(x,y) to form g(x,y)
5.3- Restoration in the Presence The local means of the pixel in Sxy
of Noise Only-Spatial Filtering
The local variance of the pixels in Sxy
5.4- Periodic Noise Reduction by
Frequency Domain Filtering
The filter can be proceeded as follows:
5.5 - Linear, Position-Invariant
Degradations The variance of the noise sh2 is zero (zero noise):
5.6- Estimating the degradation g(x,y)=f(x,y)
Function
5.7- Inverse Filtering
High local variance (edges) relative to sh2 : edges that
5.8- Minimum Mean Square should be preserved; return a value close to g(x,y)
Error (Wiener) Filtering The two variances are equal: return the arithmetic mean
5.9- Constrained Least Square value of the pixels in Sxy
Filtering
5.10- Geometric Mean Filter
5.11- Image Reconstruction from
Projections
55
Adaptive, Local Noise Reduction
Filters

5.1- A Model of the Image Purpose: Want to preserve edges

Degradation/Restoration Process
5.2- Noise Models Concept:

5.3- Restoration in the Presence 1. If sh2 is zero,  No noise

of Noise Only-Spatial Filtering the filter should return g(x,y) because g(x,y) = f(x,y)

5.4- Periodic Noise Reduction by 2. If sL2 is high relative to sh2,  Edges (should be preserved),
Frequency Domain Filtering the filter should return the value close to g(x,y)
5.5 - Linear, Position-Invariant
Degradations 3. If sL2 = sh2,  Areas inside objects
5.6- Estimating the degradation the filter should return the arithmetic mean value mL
Function
5.7- Inverse Filtering

Formula: 2

fˆ ( x, y )  g ( x, y )  2 g ( x, y )  mL 

5.8- Minimum Mean Square
Error (Wiener) Filtering
5.9- Constrained Least Square
L
Filtering
5.10- Geometric Mean Filter
5.11- Image Reconstruction from
Projections
56
Adaptive, Local Noise Reduction
Filters

5.1- A Model of the Image

Degradation/Restoration Process
5.2- Noise Models
The variance of the overall noise needs to be estimated.
5.3- Restoration in the Presence The only quantities that needs to be known or estimated:
of Noise Only-Spatial Filtering The variance of the overall noise
 2   L2
5.4- Periodic Noise Reduction by
Frequency Domain Filtering
5.5 - Linear, Position-Invariant
Degradations
5.6- Estimating the degradation
 2   L2 then the ratio (Eq. 5.3-12) is set to 1: Prevent
Function negative gray levels
5.7- Inverse Filtering
5.8- Minimum Mean Square
Error (Wiener) Filtering
5.9- Constrained Least Square
Filtering
5.10- Geometric Mean Filter
5.11- Image Reconstruction from
Projections
57
Adaptive Noise Reduction Filter -
Example

5.1- A Model of the Image

Degradation/Restoration Process
5.2- Noise Models
Image
corrupted Image
5.3- Restoration in the Presence obtained
of Noise Only-Spatial Filtering by additive
Gaussian using a 7x7
5.4- Periodic Noise Reduction by noise with arithmetic
Frequency Domain Filtering zero mean mean filter
5.5 - Linear, Position-Invariant and 2=1000
Degradations Image
5.6- Estimating the degradation obtained
Function Image
5.7- Inverse Filtering obtained using a 7x7
5.8- Minimum Mean Square using a 7x7 adaptive
Error (Wiener) Filtering geometric noise
5.9- Constrained Least Square mean filter reduction
Filtering filter
5.10- Geometric Mean Filter
5.11- Image Reconstruction from
Projections
58
Adaptive Noise Reduction Filter -
Example

5.1- A Model of the Image

Degradation/Restoration Process
5.2- Noise Models

5.3- Restoration in the Presence

of Noise Only-Spatial Filtering

5.4- Periodic Noise Reduction by

Frequency Domain Filtering
5.5 - Linear, Position-Invariant
Degradations
5.6- Estimating the degradation
Function
5.7- Inverse Filtering
5.8- Minimum Mean Square
Error (Wiener) Filtering
5.9- Constrained Least Square
Filtering
5.10- Geometric Mean Filter
5.11- Image Reconstruction from
Projections
59
Adaptive Median Filter

5.1- A Model of the Image

Degradation/Restoration Process
5.2- Noise Models

5.3- Restoration in the Presence Median filter perform well as long as spatial density of the
of Noise Only-Spatial Filtering impulse noise is not vary large (Pa and Pb <0.2 )

5.4- Periodic Noise Reduction by

Frequency Domain Filtering The adaptive filter
5.5 - Linear, Position-Invariant Can handle impulse noise with probabilities even
Degradations
larger than Pa and Pb >0.2
5.6- Estimating the degradation
Function Preserve detail while smoothing non-impulse noise
5.7- Inverse Filtering Change (increase) Sxy during filter operation
5.8- Minimum Mean Square
Error (Wiener) Filtering
5.9- Constrained Least Square
Filtering
5.10- Geometric Mean Filter
5.11- Image Reconstruction from
Projections
60
Adaptive Median Filter

5.1- A Model of the Image Purpose: Want to remove impulse noise while preserving edges
Degradation/Restoration Process
5.2- Noise Models Level A: A1= zmedian – zmin
Algorithm:
A2= zmedian – zmax
5.3- Restoration in the Presence If A1 > 0 and A2 < 0, go to level B
of Noise Only-Spatial Filtering Else increase window size
If window size <= Smax
5.4- Periodic Noise Reduction by repeat level A
Frequency Domain Filtering Else return zxy
5.5 - Linear, Position-Invariant Level B: B1= zxy – zmin
Degradations B2= zxy – zmax
5.6- Estimating the degradation If B1 > 0 and B2 < 0, return zxy
Function Else return zmedian
5.7- Inverse Filtering
5.8- Minimum Mean Square where
Error (Wiener) Filtering
5.9- Constrained Least Square zmin = minimum gray level value in Sxy
Filtering zmax = maximum gray level value in Sxy
5.10- Geometric Mean Filter zmedian = median of gray levels in Sxy
5.11- Image Reconstruction from zxy = gray level value at pixel (x,y)
Projections Smax = maximum allowed size of Sxy
61
Adaptive Median Filter: How it works

5.1- A Model of the Image

Degradation/Restoration Process
5.2- Noise Models

5.3- Restoration in the Presence

of Noise Only-Spatial Filtering Level A: A1= zmedian – zmin Determine whether
A2= zmedian – zmax zmedian is an impulse
5.4- Periodic Noise Reduction by or not
Frequency Domain Filtering
5.5 - Linear, Position-Invariant
Degradations If A1 > 0 and A2 < 0, go to level B
5.6- Estimating the degradation
Function Else  Window is not big enough
5.7- Inverse Filtering increase window size
5.8- Minimum Mean Square
Error (Wiener) Filtering If window size <= Smax
5.9- Constrained Least Square repeat level A
Filtering
5.10- Geometric Mean Filter
Else
5.11- Image Reconstruction from return zxy
Projections
62
Adaptive Median Filter: How it works

5.1- A Model of the Image

Degradation/Restoration Process
5.2- Noise Models

5.3- Restoration in the Presence Level B:  zmedian is not an impulse

of Noise Only-Spatial Filtering Determine whether zxy
is an impulse or not
5.4- Periodic Noise Reduction by
B1= zxy – zmin
Frequency Domain Filtering B2= zxy – zmax
5.5 - Linear, Position-Invariant
Degradations
5.6- Estimating the degradation If B1 > 0 and B2 < 0,  zxy is not an
Function impulse
5.7- Inverse Filtering
5.8- Minimum Mean Square
return zxy  to preserve
Error (Wiener) Filtering original details
5.9- Constrained Least Square Else
Filtering
5.10- Geometric Mean Filter
return zmedian  to remove
5.11- Image Reconstruction from impulse
Projections
63
Adaptive Median Filter - Example

5.1- A Model of the Image

Degradation/Restoration Process
5.2- Noise Models

5.3- Restoration in the Presence

of Noise Only-Spatial Filtering

5.4- Periodic Noise Reduction by

Frequency Domain Filtering
5.5 - Linear, Position-Invariant
Degradations Image corrupted Image obtained Image obtained
5.6- Estimating the degradation by salt-and-pepper using a 7x7 using an adaptive
Function noise with median filter median filter with
5.7- Inverse Filtering pa=pb= 0.25 Smax = 7
5.8- Minimum Mean Square
Error (Wiener) Filtering
5.9- Constrained Least Square More small details are preserved
Filtering
5.10- Geometric Mean Filter
5.11- Image Reconstruction from
Projections
64
5.4 Periodic Noise Reduction by
Frequency Domain Filtering

65
Periodic Noise Reduction by Frequency
Domain Filtering

5.1- A Model of the Image

Degradation/Restoration Process
5.2- Noise Models
5.3- Restoration in the Presence
of Noise Only-Spatial Filtering Bandreject Filters
5.4- Periodic Noise Reduction by
Frequency Domain Filtering
Bandpass Filters
5.5 - Linear, Position-Invariant
Degradations
5.6- Estimating the degradation Notch Filters
Function
5.7- Inverse Filtering
5.8- Minimum Mean Square
Error (Wiener) Filtering Optimum Notch Filtering
5.9- Constrained Least Square
Filtering
5.10- Geometric Mean Filter
5.11- Image Reconstruction from
Projections
66
Bandreject Filters

5.1- A Model of the Image

Degradation/Restoration Process
5.2- Noise Models  Use to eliminate frequency components in some bands.
5.3- Restoration in the Presence  Ideal Band-reject Filter:
of Noise Only-Spatial Filtering

5.4- Periodic Noise Reduction by -D(u,v) =distance from the origin H (u, v)  1 if D(u, v)  D0 
W
Frequency Domain Filtering 2
of the centered freq. rectangle W W
0 if D0   D(u, v)  D0 
5.5 - Linear, Position-Invariant -W =width of the band 2 2
W
Degradations -D0=Radial center of the band. 1 if D(u, v)  D0 
2
5.6- Estimating the degradation
Function
5.7- Inverse Filtering
5.8- Minimum Mean Square
Error (Wiener) Filtering
5.9- Constrained Least Square
Filtering
5.10- Geometric Mean Filter
5.11- Image Reconstruction from
Projections
67
Bandreject Filters

5.1- A Model of the Image

Degradation/Restoration Process
5.2- Noise Models Degraded image DFT
5.3- Restoration in the Presence
of Noise Only-Spatial Filtering

5.4- Periodic Noise Reduction by

Frequency Domain Filtering
Periodic noise
5.5 - Linear, Position-Invariant
Degradations can be reduced by
5.6- Estimating the degradation setting frequency
Function components
5.7- Inverse Filtering corresponding to
5.8- Minimum Mean Square noise to zero.
Error (Wiener) Filtering
5.9- Constrained Least Square
Filtering
5.10- Geometric Mean Filter Bandreject filter Restored image
5.11- Image Reconstruction from
Projections
68
Bandpass Filters

5.1- A Model of the Image

Degradation/Restoration Process Opposite operation of a band-reject filter:
5.2- Noise Models
5.3- Restoration in the Presence
of Noise Only-Spatial Filtering H bp  1  H br (u , v)
5.4- Periodic Noise Reduction by
Frequency Domain Filtering

5.5 - Linear, Position-Invariant

Degradations
5.6- Estimating the degradation
Function
5.7- Inverse Filtering
5.8- Minimum Mean Square
Error (Wiener) Filtering
5.9- Constrained Least Square
Filtering
5.10- Geometric Mean Filter
5.11- Image Reconstruction from Periodic noise from the previous slide that is Filtered out.
Projections
69
Notch Filters

5.1- A Model of the Image  A notch reject filter is used to eliminate some frequency
Degradation/Restoration Process components.
5.2- Noise Models
5.3- Restoration in the Presence  Rejects (or passes) frequencies in predefined neighborhoods
of Noise Only-Spatial Filtering about a center frequency.
Ideal
5.4- Periodic Noise Reduction by
Frequency Domain Filtering Must
5.5 - Linear, Position-Invariant
appear in
Degradations symmetri
5.6- Estimating the degradation c pairs
Function
5.7- Inverse Filtering
about the
5.8- Minimum Mean Square origin.
Error (Wiener) Filtering
5.9- Constrained Least Square Butterworth Gaussian
Filtering
5.10- Geometric Mean Filter
5.11- Image Reconstruction from
Projections
70
Notch reject Filter - Example

5.1- A Model of the Image Degraded image Notch filter

Degradation/Restoration Process DFT (freq. Domain)
5.2- Noise Models
5.3- Restoration in the Presence
of Noise Only-Spatial Filtering

5.4- Periodic Noise Reduction by

Frequency Domain Filtering

5.5 - Linear, Position-Invariant

Degradations
5.6- Estimating the degradation
Function
5.7- Inverse Filtering
5.8- Minimum Mean Square
Error (Wiener) Filtering
5.9- Constrained Least Square
Filtering
5.10- Geometric Mean Filter
5.11- Image Reconstruction from
Projections
71 Noise Restored image
Notch reject Filter - Example

5.1- A Model of the Image

Degradation/Restoration Process
5.2- Noise Models
5.3- Restoration in the Presence
of Noise Only-Spatial Filtering

5.4- Periodic Noise Reduction by

Frequency Domain Filtering

5.5 - Linear, Position-Invariant

Degradations
5.6- Estimating the degradation
Function
5.7- Inverse Filtering
5.8- Minimum Mean Square
Error (Wiener) Filtering
5.9- Constrained Least Square
Filtering
5.10- Geometric Mean Filter
5.11- Image Reconstruction from
Projections
72
Notch reject Filter - Example

5.1- A Model of the Image

Degradation/Restoration Process
5.2- Noise Models
5.3- Restoration in the Presence
of Noise Only-Spatial Filtering

5.4- Periodic Noise Reduction by

Frequency Domain Filtering

5.5 - Linear, Position-Invariant

Degradations
5.6- Estimating the degradation
Function
5.7- Inverse Filtering
5.8- Minimum Mean Square
Error (Wiener) Filtering
5.9- Constrained Least Square
Filtering
5.10- Geometric Mean Filter
5.11- Image Reconstruction from
Projections
73
 Image Degraded by Periodic Noise

Degraded
DFT (no shift)
image
5.1- A Model of the Image
Degradation/Restoration Process
5.2- Noise Models
5.3- Restoration in the Presence
of Noise Only-Spatial Filtering

5.4- Periodic Noise Reduction by

Frequency Domain Filtering

5.5 - Linear, Position-Invariant Several pairs of components are present  more than just one
Degradations
5.6- Estimating the degradation sinusoidal component
Function
5.7- Inverse Filtering
5.8- Minimum Mean Square
Error (Wiener) Filtering
5.9- Constrained Least Square
Filtering
5.10- Geometric Mean Filter
5.11- Image Reconstruction from
Projections DFT of noise Noise Restored image
74
5.5 Linear, Position-Invariant
Degradations

75
 Linear, Position-invarient
Degradations
• The input-output relationship before the restoration
stage :
g ( x, y )  H [ f ( x, y )]   ( x, y )
Let us assume that :

 ( x, y )  0
 So that :

g ( x, y )  H [ f ( x, y )]
76
Linear, Position-invarient
Degradations–cont
• H is linear if
H [af 1( x, y )  bf 2( x, y )]  aH [ f 1( x, y )]  bH [ f 2( x, y )]

 An operator having the input-output relationship

is said to be position(space) invarient if
H [ f ( x   , y   )]  g ( x   , y   )
For any f(x,y) and any α and β.

77
 Linear, Position-invarient
Degradations–cont
 f(x,y) can be expressed in terms of a continuous impulse
function:

 
f ( x, y )    f ( ,  ) ( x   , y   )dd
 


g ( x, y )  H [ f ( x, y )]  H 
 
f ( ,  ) ( x   , y   )dd 
    

f ( ,  ) H  ( x   , y   )dd
 
g ( x, y )   
 

H  ( x   , y   )  h( x,  , y,  )
 
g ( x, y )   
f ( ,  )h( x,  , y,  )78
dd
Linear, Position-invarient
Degradations–cont
 If H is position invarient, then
H  ( x   , y   )  h( x   , y   )

 
g ( x, y )    f ( ,  )h( x   , y   )dd
 

 
g ( x, y )    f ( ,  )h( x   , y   )dd   ( x, y )
 

g ( x, y )  h ( x, y )  f ( x, y )   ( x, y )
G (u , v)  H (u , v) F (u , v)  N (u , v)
79
Linear, Position-invarient
Degradations–cont

• Many types of degradation can be approximated by linear,

position-invariant processes
•In this situation, restoration is image deconvolution

g ( x, y )  h ( x, y )  f ( x, y )   ( x, y )
G (u , v)  H (u , v) F (u , v)  N (u , v)
80
 5.6 Estimating the degradation
Function

81
Estimating the Degradation Function

5.1- A Model of the Image

Degradation/Restoration Process
5.2- Noise Models
5.3- Restoration in the Presence
of Noise Only-Spatial Filtering Estimation by Image Observation
5.4- Periodic Noise Reduction by
Frequency Domain Filtering
5.5 - Linear, Position-Invariant
Degradations Estimation by Experimentation

5.6- Estimating the degradation

Function
Estimation by Modeling
5.7- Inverse Filtering
5.8- Minimum Mean Square
Error (Wiener) Filtering
5.9- Constrained Least Square
Filtering
5.10- Geometric Mean Filter
5.11- Image Reconstruction from
Projections
82
Estimating the Degradation Function

Degradation model:

g ( x, y )  f ( x, y )  h( x, y )   ( x, y )
5.1- A Model of the Image
Degradation/Restoration Process
5.2- Noise Models
5.3- Restoration in the Presence
or
of Noise Only-Spatial Filtering
5.4- Periodic Noise Reduction by
G ( u, v )  F ( u, v ) H ( u, v )  N ( u, v )
Frequency Domain Filtering
5.5 - Linear, Position-Invariant Purpose: To estimate h(x,y) or H(u,v)
Degradations

5.6- Estimating the degradation Why? If we know exactly h(x,y), regardless of noise, we can do
Function
deconvolution to get f(x,y) back from g(x,y).
5.7- Inverse Filtering
5.8- Minimum Mean Square
Error (Wiener) Filtering
Methods:
5.9- Constrained Least Square
Filtering 1. Estimation by Image Observation
5.10- Geometric Mean Filter
5.11- Image Reconstruction from 2. Estimation by Experiment
Projections
83
3. Estimation by Modeling
Estimation by Image Observation

5.1- A Model of the Image

Degradation/Restoration Process
5.2- Noise Models
5.3- Restoration in the Presence
of Noise Only-Spatial Filtering
5.4- Periodic Noise Reduction by
Frequency Domain Filtering
5.5 - Linear, Position-Invariant
Degradations

5.6- Estimating the degradation

Function

5.7- Inverse Filtering

5.8- Minimum Mean Square
Error (Wiener) Filtering
5.9- Constrained Least Square
Filtering
5.10- Geometric Mean Filter
5.11- Image Reconstruction from
Projections
84
Estimation by Image Observation

Original image (unknown) Degraded image

g(x,y)
5.1- A Model of the Image
Degradation/Restoration Process
f(x,y)*h(x,y)
5.2- Noise Models
5.3- Restoration in the Presence
of Noise Only-Spatial Filtering
5.4- Periodic Noise Reduction by
Frequency Domain Filtering f(x,y) Observation
5.5 - Linear, Position-Invariant Subimage
Degradations DFT
Gs ( u , v )
5.6- Estimating the degradation Estimated Transfer g s ( x, y )
Function
function
Gs (u, v )
5.7- Inverse Filtering H ( u, v )  H s ( u, v )  Restoration
5.8- Minimum Mean Square Fˆs (u, v ) process by
estimation
Error (Wiener) Filtering
5.9- Constrained Least Square
DFT
Fˆs (u, v )
Filtering
5.10- Geometric Mean Filter This case is used when we Reconstructed
5.11- Image Reconstruction from know only g(x,y) and cannot Subimage
Projections
85 repeat the experiment! fˆs ( x, y )
Estimation by Experimentation
Estimation by Experiment

5.1- A Model of the Image

Degradation/Restoration Process
5.2- Noise Models
5.3- Restoration in the Presence
of Noise Only-Spatial Filtering
5.4- Periodic Noise Reduction by
Frequency Domain Filtering
5.5 - Linear, Position-Invariant
Degradations

5.6- Estimating the degradation

Function

5.7- Inverse Filtering

5.8- Minimum Mean Square
Error (Wiener) Filtering
5.9- Constrained Least Square
Filtering
5.10- Geometric Mean Filter
5.11- Image Reconstruction from
Projections
86
Estimation by Experimentation
Estimation by Experiment

o Used when we have the same equipment set up and can repeat
5.1- A Model of the Image the experiment.
Degradation/Restoration Process Response image from
5.2- Noise Models Input impulse image the system
5.3- Restoration in the Presence
of Noise Only-Spatial Filtering
5.4- Periodic Noise Reduction by
Frequency Domain Filtering
5.5 - Linear, Position-Invariant System
Degradations
H( )
5.6- Estimating the degradation
Function
A ( x , y ) g ( x, y )
5.7- Inverse Filtering
5.8- Minimum Mean Square DFT DFT
Error (Wiener) Filtering
5.9- Constrained Least Square
Filtering DFT A ( x, y )  A G (u, v )
5.10- Geometric Mean Filter
G ( u, v )
5.11- Image Reconstruction from
Projections H ( u, v ) 
87 A
Estimation by Modeling

5.1- A Model of the Image

Degradation/Restoration Process
5.2- Noise Models
5.3- Restoration in the Presence
of Noise Only-Spatial Filtering
5.4- Periodic Noise Reduction by
Frequency Domain Filtering
5.5 - Linear, Position-Invariant
Degradations

5.6- Estimating the degradation

Function

5.7- Inverse Filtering

5.8- Minimum Mean Square
Error (Wiener) Filtering K is a parameter to be determined by experiments
5.9- Constrained Least Square because it changes with the nature of turbulence
Filtering
5.10- Geometric Mean Filter
5.11- Image Reconstruction from
Projections
88
Estimation by Modeling

5.1- A Model of the Image

Degradation/Restoration Process  Used when we know physical mechanism underlying the
5.2- Noise Models image formation process that can be expressed mathematically.
5.3- Restoration in the Presence
of Noise Only-Spatial Filtering
Original image Severe turbulence Example:
5.4- Periodic Noise Reduction by
Frequency Domain Filtering Atmospheric
5.5 - Linear, Position-Invariant Turbulence model
Degradations

5.6- Estimating the degradation k = 0.0025

Function
 k ( u 2  v 2 )5 / 6
5.7- Inverse Filtering Mild turbulence Low turbulence H ( u, v )  e
5.8- Minimum Mean Square
Error (Wiener) Filtering
5.9- Constrained Least Square
Filtering
5.10- Geometric Mean Filter
5.11- Image Reconstruction from k = 0.001 k = 0.00025
Projections
89
Estimation by Modeling: Motion Blurring

Suppose that an image f(x,y) undergoes planar motion and

that x0(t) and y0(t) are the time varying components of motion
5.1- A Model of the Image
Degradation/Restoration Process ( x0 (t ), y0 (t ))
5.2- Noise Models
5.3- Restoration in the Presence The total exposure at any point of the recording medium is
of Noise Only-Spatial Filtering obtained by integrating the instantaneous exposure over the time
5.4- Periodic Noise Reduction by interval(T) during which the imaging system shutter is open.
Frequency Domain Filtering
5.5 - Linear, Position-Invariant The blurred image is obtained by
Degradations T
g ( x, y )   f ( x  x0 (t ), y  y0 (t ))dt
5.6- Estimating the degradation 0
Function
where T = exposure time.
 

  g ( x, y )e
 j 2 ( ux  vy )
5.7- Inverse Filtering G (u, v )  dxdy
5.8- Minimum Mean Square   
Error (Wiener) Filtering  
T 
5.9- Constrained Least Square      f ( x  x0 (t ), y  y0 (t ))dt  e  j 2 ( ux  vy ) dxdy
Filtering
    0 
5.10- Geometric Mean Filter
5.11- Image Reconstruction from  
T

Projections      f ( x  x0 (t ), y  y0 (t ))e  j 2 ( ux  vy) dxdy dt
90 0     
Estimation by Modeling: Motion Blurring

G (u , v)   F (u , v)e  j 2 ux 0 ( t )  vy 0 ( t )  dt
T
 F (u , v)  e  j 2 ux 0 ( t )  vy 0 ( t )  dt
T

0 0

 j 2 ux 0 ( t )  vy 0 ( t ) 
T
H (u , v )   0
e dt

G (u , v)  H (u , v) F (u , v)
If the motion variables x0(t)and y0(t) are known, the transfer function H(u,v) can
be obtained directly from above equation

91
Motion Blurring - Example

5.1- A Model of the Image For constant motion x0(t)=at/T y0(t)=bt/T

Degradation/Restoration Process
T
5.2- Noise Models
H ( u, v )  sin( (ua  vb))e  j ( ua vb)
5.3- Restoration in the Presence
of Noise Only-Spatial Filtering
 (ua  vb)
5.4- Periodic Noise Reduction by
Frequency Domain Filtering
5.5 - Linear, Position-Invariant
Degradations

5.6- Estimating the degradation

Function

5.7- Inverse Filtering

5.8- Minimum Mean Square
Error (Wiener) Filtering
5.9- Constrained Least Square Original image Motion blurred image
Filtering a = b = 0.1, T = 1
5.10- Geometric Mean Filter
5.11- Image Reconstruction from
Projections
92
PSF (Point Spread Function)

5.1- A Model of the Image

Degradation/Restoration Process
5.2- Noise Models
5.3- Restoration in the Presence
of Noise Only-Spatial Filtering 2D Equivalent to Impulse Response
5.4- Periodic Noise Reduction by
Frequency Domain Filtering What happen to a single point of light when it passes
5.5 - Linear, Position-Invariant through a system?
Degradations
PSF describes a LSI system
5.6- Estimating the degradation In practise PSF should be estimated
Function

5.7- Inverse Filtering

5.8- Minimum Mean Square
Error (Wiener) Filtering
5.9- Constrained Least Square
Filtering
5.10- Geometric Mean Filter
5.11- Image Reconstruction from
Projections
93
 5.7 Inverse Filtering

94
Inverse Filtering

5.1- A Model of the Image

Degradation/Restoration Process
5.2- Noise Models
5.3- Restoration in the Presence
Based on properties of the Fourier transforms
of Noise Only-Spatial Filtering Assume degradation can be expressed as convolution
5.4- Periodic Noise Reduction by
Frequency Domain Filtering g ( x, y )  f ( x, y )   ( x, y )
5.5 - Linear, Position-Invariant
Degradations
After applying the Fourier transform to Eq. (a), we get
5.6- Estimating the degradation
Function
G (u , v)  H (u , v) F (u , v)  N (u , v)
5.7- Inverse Filtering
An estimate Fˆ(u,v) of the transform of the original
image
5.8- Minimum Mean Square
Error (Wiener) Filtering G (u, v)
5.9- Constrained Least Square Fˆ (u, v) 
Filtering H (u, v)
5.10- Geometric Mean Filter
5.11- Image Reconstruction from
Projections
95
Inverse Filtering

5.1- A Model of the Image

Degradation/Restoration Process
5.2- Noise Models
5.3- Restoration in the Presence
of Noise Only-Spatial Filtering
5.4- Periodic Noise Reduction by
Frequency Domain Filtering
5.5 - Linear, Position-Invariant
Degradations
5.6- Estimating the degradation
Function

5.7- Inverse Filtering

5.8- Minimum Mean Square

Error (Wiener) Filtering
5.9- Constrained Least Square
Filtering
5.10- Geometric Mean Filter
5.11- Image Reconstruction from
Projections
96
Inverse Filtering

5/6
5.1- A Model of the Image  k ( u  M / 2) 2  ( v  N / 2) 2 
Degradation/Restoration Process Degradation function H (u , v)  e
5.2- Noise Models
5.3- Restoration in the Presence
of Noise Only-Spatial Filtering
5.4- Periodic Noise Reduction by
Frequency Domain Filtering
5.5 - Linear, Position-Invariant
Degradations
5.6- Estimating the degradation
Function

5.7- Inverse Filtering

5.8- Minimum Mean Square

Error (Wiener) Filtering
5.9- Constrained Least Square
Filtering
5.10- Geometric Mean Filter
5.11- Image Reconstruction from
Projections
Cutting off values of the ratio outside a radius of 40, 70,85.
97
Restoration Cut-off Frequency

5.1- A Model of the Image

Degradation/Restoration Process
5.2- Noise Models
5.3- Restoration in the Presence
of Noise Only-Spatial Filtering Limiting the restoration to a specific frequency
5.4- Periodic Noise Reduction by about the origin
Frequency Domain Filtering
5.5 - Linear, Position-Invariant
Degradations
5.6- Estimating the degradation Result:
Function
• Low-pass image
5.7- Inverse Filtering • Blurred
5.8- Minimum Mean Square • Ringing
Error (Wiener) Filtering
5.9- Constrained Least Square
Filtering
5.10- Geometric Mean Filter
5.11- Image Reconstruction from
Projections
98
Inverse Filtering - Example

5.1- A Model of the Image

Degradation/Restoration Process
5.2- Noise Models
5.3- Restoration in the Presence
of Noise Only-Spatial Filtering
5.4- Periodic Noise Reduction by
Frequency Domain Filtering
5.5 - Linear, Position-Invariant
Degradations
5.6- Estimating the degradation
Function

5.7- Inverse Filtering

5.8- Minimum Mean Square

Error (Wiener) Filtering
5.9- Constrained Least Square
Filtering
5.10- Geometric Mean Filter
5.11- Image Reconstruction from
Projections
99
 5.8 Minimum Mean Square Error
(Wiener) Filtering

101
Minimum Mean Square Error (Wiener)
Filtering

5.1- A Model of the Image

Degradation/Restoration Process
5.2- Noise Models
5.3- Restoration in the Presence
of Noise Only-Spatial Filtering
5.4- Periodic Noise Reduction by
Frequency Domain Filtering
5.5 - Linear, Position-Invariant
Degradations
5.6- Estimating the degradation
Function
5.7- Inverse Filtering

5.8- Minimum Mean Square

Error (Wiener) Filtering

5.9- Constrained Least Square

Filtering
5.10- Geometric Mean Filter
5.11- Image Reconstruction from
Projections
102
Minimum Mean Square Error (Wiener)
Filtering

5.1- A Model of the Image

Degradation/Restoration Process The inverse filtration gives poor results, since the
5.2- Noise Models information about noise properties is not taken into
5.3- Restoration in the Presence account.
of Noise Only-Spatial Filtering
5.4- Periodic Noise Reduction by Wiener filtration incorporates a priori knowledge about the
Frequency Domain Filtering noise properties.
5.5 - Linear, Position-Invariant
Degradations Restoration by the filter gives an estimate f of the original
5.6- Estimating the degradation uncorrupted image f with minimal mean square error :
Function Minimization is easy if the estimate f is a linear
5.7- Inverse Filtering combination of the values in the image g;
5.8- Minimum Mean Square The estimate F of the Fourier transform of the
Error (Wiener) Filtering original image f can be expressed as:

Fˆ (u , v)  H W (u , v)G (u , v)
5.9- Constrained Least Square
Filtering
5.10- Geometric Mean Filter
5.11- Image Reconstruction from
Projections
103
Minimum Mean Square Error (Wiener)
Filtering

5.1- A Model of the Image

Degradation/Restoration Process
5.2- Noise Models
5.3- Restoration in the Presence
of Noise Only-Spatial Filtering
5.4- Periodic Noise Reduction by
Frequency Domain Filtering
5.5 - Linear, Position-Invariant
Degradations
5.6- Estimating the degradation
Function
5.7- Inverse Filtering

5.8- Minimum Mean Square

Error (Wiener) Filtering

5.9- Constrained Least Square

Filtering
5.10- Geometric Mean Filter
5.11- Image Reconstruction from
Projections
104
Minimum Mean Square Error (Wiener)
Filtering

5.1- A Model of the Image

Degradation/Restoration Process
5.2- Noise Models
5.3- Restoration in the Presence
of Noise Only-Spatial Filtering
5.4- Periodic Noise Reduction by
Frequency Domain Filtering
5.5 - Linear, Position-Invariant
Degradations
5.6- Estimating the degradation
Function
5.7- Inverse Filtering

5.8- Minimum Mean Square

Error (Wiener) Filtering

5.9- Constrained Least Square

Filtering
5.10- Geometric Mean Filter
5.11- Image Reconstruction from
Projections
105
Inverse & Wiener Filtering - Example

5.1- A Model of the Image

Degradation/Restoration Process
5.2- Noise Models
5.3- Restoration in the Presence
of Noise Only-Spatial Filtering
5.4- Periodic Noise Reduction by
Frequency Domain Filtering
5.5 - Linear, Position-Invariant
Degradations
5.6- Estimating the degradation
Function
5.7- Inverse Filtering

5.8- Minimum Mean Square

Error (Wiener) Filtering

5.9- Constrained Least Square

Filtering
5.10- Geometric Mean Filter
5.11- Image Reconstruction from
Projections
106
Inverse & Wiener Filtering -Example

5.1- A Model of the Image

Degradation/Restoration Process
5.2- Noise Models
5.3- Restoration in the Presence
of Noise Only-Spatial Filtering
5.4- Periodic Noise Reduction by
Frequency Domain Filtering
5.5 - Linear, Position-Invariant
Degradations
5.6- Estimating the degradation
Function
5.7- Inverse Filtering

5.8- Minimum Mean Square

Error (Wiener) Filtering

5.9- Constrained Least Square

Filtering
5.10- Geometric Mean Filter
5.11- Image Reconstruction from
Projections
107
Inverse & Wiener Filtering -Example

K=0.1 K=0.01 K=0.001

108
 signal-to-noise ratio in the
frequncy domain
 One of most important is the signal-to-noise ratio,
approximated using frequency domain quantities such
as M 1 N 1

 F (u, v)
2

SNR  u 0 v 0
M 1 N 1

 N (u, v)
2

u 0 v 0

• This ratio gives of the level of information bearing

signal power to the level of noise power
• Images with low noise tend to have a high SNR and,
conversely, the same image with a higher level of noise
has a lower SNR 109
Mean Square Error
 Mean square error given in statistical form in

 ˆ 2
e  E ( f  f ) 
2

 
 Can be approximated in terms a summation involving
the original and restored image:

 ˆ ( x, y ) 
M 1 N 1
1

2
MSE  f ( x , y )  f
MN x 0 y 0

110
 signal-to-noise ratio in the
spatial domain
 If one considers the restored image to be “signal” and
the difference between this image and the orginal to be
noise, we can define a signal-to-noise ratio in the
spatial domain
M 1 N 1


x 0 y 0
fˆ ( x, y ) 2
SNR 
  
M 1 N 1
ˆ
f ( x, y )  f ( x, y )
2

x 0 y 0

• The closer f and f̂ are, the larger this ratio will be.
111
 5.9 Constrained Least Square
Filtering

112
Constrained Least Squares Filtering

5.1- A Model of the Image Provides a filter that can eliminate some of the artifacts caused by
Degradation/Restoration Process other frequency domain filters
5.2- Noise Models
5.3- Restoration in the Presence
of Noise Only-Spatial Filtering Done by smoothing criterion in the filter derivation
5.4- Periodic Noise Reduction by
Frequency Domain Filtering
5.5 - Linear, Position-Invariant The result does not have undesirable oscillations
Degradations
5.6- Estimating the degradation
Function   H * (u , v) 
5.7- Inverse Filtering F (u , v)   G (u , v)
 H (u , v)   P (u , v) 
2 2
5.8- Minimum Mean Square
Error (Wiener) Filtering
P(u,v) is the fourier transform of the Laplacian operator
5.9- Constrained Least Square
Filtering

5.10- Geometric Mean Filter

5.11- Image Reconstruction from
Projections
113
Constrained Least Squares Filtering

 An additional difficult in Wiener Filter:

 The power spectra of the undergraded image and
noise must be known.

 A constant estimate of the ratio of the power

spectra is not always a suitable solution.
114
Constrained Least Squares Filtering

Properties
of CLSF:
Prior knowledge
 Require knowledge of only mean and variance of the noise.
 Can be calculated from a given degraded image
Difference in Criterion
 Wiener filter: statistically optimal in an average sense.
 New approach: optimal for each image.
Not related to the dynamics of visual perception.
The choice of algorithm is determined by the perceive
visual quality of the resulting images.

115
Constrained Least Squares Filtering

Find the minimum of a criterion function C:

 
M 1 N 1
C    f ( x, y ) 2 2

x 0 y 0
Subject to the constraint: ˆ
|| g  Hf || || η ||
2 2

ˆ  H *
(u, v) 
F (u, v)   2
G(u, v)
 | H (u, v) |  | P(u, v) | 
2

 0 －1 0 
p ( x, y )＝ －1 4 －1
 0 －1 0 
𝛾 : adjustable parameter
𝑃(𝑢, 𝑣) : Fourier transform of the Laplacian operator
Constrained Least Squares Filtering

 Adjust 𝛾 interactively until acceptable results are

achieved.
 Optimality, adjust 𝛾 until the constraint is satisfied.

117
Procedure for computing

118
 Constrained Least Squares Filtering
Gaussian noise:
mean=0,
variance=650

Variance=65

Variance=0.0065

119
Constrained Least Squares Filtering

 Remarks:
 Optimum restoration does not necessarily imply “best”
in the visual sense.
 Depending on the nature and magnitude of the
degradation and noise, it is best to iteratively
determining the optimum parameters.
 Usually, automatic determined filters yield inferior
results to manual adjustment of filter parameters.
Especially for a single, scalar parameter in CLSF.

120
Constrained Least Squares Filtering

5.1- A Model of the Image

Degradation/Restoration Process
5.2- Noise Models
5.3- Restoration in the Presence
of Noise Only-Spatial Filtering
5.4- Periodic Noise Reduction by
Frequency Domain Filtering
5.5 - Linear, Position-Invariant
Degradations
5.6- Estimating the degradation
Function
5.7- Inverse Filtering
5.8- Minimum Mean Square
Error (Wiener) Filtering

5.9- Constrained Least Square

Filtering

5.10- Geometric Mean Filter

5.11- Image Reconstruction from
Projections
121
 5.10 Geometric Mean filter

122
Geometric Mean Filter

5.1- A Model of the Image

Degradation/Restoration Process
5.2- Noise Models 1
5.3- Restoration in the Presence  
of Noise Only-Spatial Filtering  
 H * (u , v)   H *
(u , v ) 
Fˆ (u , v)  
5.4- Periodic Noise Reduction by
  G (u , v)
 S (u , v)  
Frequency Domain Filtering 2
 H (u , v)  
H (u , v)    
5.5 - Linear, Position-Invariant 2
Degradations   S f (u , v)  
5.6- Estimating the degradation 
Function

 1
5.7- Inverse Filtering
5.8- Minimum Mean Square
Error (Wiener) Filtering
5.9- Constrained Least Square
Filtering
 0
5.10- Geometric Mean Filter
 1
5.11- Image Reconstruction from
Projections
123