Introduction

 The main objective of restoration is to improve the quality of a digital

image which has been degraded due to Various phenomena like:
• Motion
• Improper focusing of Camera during image acquisition.
• Noise

 The purpose of image restoration is to restore a degraded/distorted

image to its original content and quality. It is based on mathematical
models of image degradation.

 Restoration involves following process:-

• Modeling of Degradation
• Applying the inverse process to recover the original image
Image Enhancement vs. Image Restoration

 Image enhancement: process image so that the result is more suitable

for a specific application, is largely a subjective process.

 Image restoration: recover image from distortions to its original image,

is largely an objective process.
S. No. Image Enhancement Image Restoration
The output image looks
The output image is enhanced
1 perfect, as if it was taken
for specific application.
without any flaws or damage.
It makes image look better
It tries to fix the image to get
2 without regard to how it truly
back to the real, true image.
should look.
This means improving to show This means improving image to
3
some hidden details. match the original image.
4 This purely subjective. This is purely objective.
This is a cosmetic procedure This is a reverse procedure of
5 of improving quality by getting an optimal estimate of
working with existing data . the desired result.

ASRao 5
Properties of Noise

 A noise is a white noise when its Fourier Spectrum is constant.

• We assume that the noise is independent of the spatial coordinates, and

that it is uncorrelated with respect to image itself.

• Noise cannot be predicted, but can be approximately described in

statistical way using the probability density function (PDF).

• Noise is very difficult to remove it from the digital images without the
prior knowledge of noise model.

• So, noise models are essential in the study of image denoising

techniques.
Noise models
Noise is always presents in digital images during image acquisition, coding,
transmission and processing steps.
First we consider the degradation due to noise only
 Source of noise
 Image acquisition
 Image transmission

 In Acquiring images with Camera and Light levels are major

factor affecting the amount of noise in resulting image.

 Images are corrupted during transmission due to interferences in

the channel used for transmission.
Example: Image is transferred using a wireless network might be
corrupted due to atmospheric disturbances.
Consider a noisy image modelled as

g ( x, y )  f ( x, y )   ( x, y )

where
f(x, y) - Original Image η(x, y) – Noise g(x,y) – Noisy image

 Spatial properties of noise

 Statistical behavior of the gray-level values of pixels
 Noise parameters, correlation with the image.

 Frequency properties of noise

 Fourier spectrum
Example: White noise (a constant
ASRao Fourier spectrum 8
White noise
 Autocorrelation function is an impulse function multiplied by a
constant.
 It means there is no correlation between any two pixels in the
noise image.
 There is no way to predict the next noise value.
The spectrum of the autocorrelation function is a constant (white)
Types of Noise Models

 Gaussian Noise (most common model)

 Rayleigh Noise

 Erlang (Gamma) Noise

 Uniform Noise

 Exponential Noise

 Salt and Pepper Noise

Noise Probability Density Functions

 Noises are taken as random variables.

 Noise cannot be predicated but can be approximately described in

statistical way using the probability density function (PDF)

 PDF tells how much each value occurs.

 Noise (image) can be classified according the distribution of the values

of pixels (of the noise image) or its (normalized) histogram.

 Most common PDFs found in image processing applications:

- Gaussian noise - Impulse (Salt & Pepper) noise
- Rayleigh noise
- Erlang (Gamma) noise
- Uniform noise
Gaussian Noise

 Gaussian noise is characterized by two parameters, μ (mean) and σ2

(variance) by

1 ( z   ) 2 / 2 2
p( z )  e
2 

 70% of its value will be in the range of

   ,    
Rayleigh Noise
Erlang (Gamma) Noise
Exponential Noise

 Exponential PDF is a special case of Erlang PDF with b=1.

Uniform Noise
Impulse (Salt & Pepper) Noise

 It is the only type of noise that can be distinguished fromm others

visually.
Sources of Noise

 Gaussian noise arises in an image due to factors such as electronic

circuit noise and sensor noise due to poor illumination and/or high
temperature.

 Raleigh noise is helpful in characterizing noise phenomena in range

imaging.

 Exponential and Gamma noises are more common is Laser Imaging.

 Impulse noise is found in situations where quick transients, such as

faulty switching take place during imaging.
Effect of Noise on Images & Histograms

 The test pattern shown here is ideal for demonstrating the addition of
noise.
Image Degradation with Additive Noise
ASRao 22
Model for Image Degradation/Restoration Process

 We will assume that a degradation function (h(x, y)) exists, which,

together with additive noise (η(x, y)), operates on the input image
f(x,y) to produce a degraded image g(x,y).

 The objective of restoration is to obtain an estimate for the original

image from its degraded version g(x,y) while having some knowledge
about the degradation function H and the noise.
 The degraded image in the spatial domain is

where h(x,y) is the spatial representation of the degradation

function and η(x,y) is the noise

 The degraded image in the frequency domain is

Linear, Position-Invariant Degradation
 From the model of image degradation, one may write:

where H is the degradation function and η is the additive noise

 Initially assume that η(x,y) = 0 which gives,

 An operator having input-output relationship g(x,y) = H[f(x,y)] is said to

be position-(or space-) invariant if

for any f(x,y) and any α, β

ASRao 25
 The definition indicates that the response at any point in the image
depends only on the value of the input at that point, not on its position.

 The following function is called a Point Spread Function:

the above is nothing but the impulse response of H

 Now we may express the linear position-invariant degradation (Spatial

and Frequency Domain) as follows:

ASRao 26
Estimation of Degradation Model

 Degradation comes in many forms such as motion blur, noise and camera
misfocus.

 In cases like motion blur, it is possible to come up with an very good

estimate of the actual blurring function and “undo” the blur to restore
the original image.

 Degradation model:

 Purpose: to estimate h(x, y) or H(u, v)

 Why: if we know h(x, y) regardless of noise, we can do deconvolution to

get f(x, y) back form g(x, y) ASRao 27
 Degradation function is important in all cases of knowledge like
frequency domain, spatial domain or matrix.

 Estimation of degradation is important.

 There are three ways to estimate:

 By Observation
 By Experimentation
 By Mathematical Modeling

 The degradations are modelled as being result of convolution, and

restoration seeks to find filters that apply process in reverse.

 Due to that reason, the term image deconvolution is used frequently to

signify linear image restoration.
 The filters used in the restoration process often are called
deconvolution function.

 It is due to the fact that true degradation function is seldom known

completely.

ASRao 29
Estimation by Observation
 Let the image is degraded by an unknown degradation function H.

 Assume that H is linear and position invariant.

 Gather information from the image itself.

 We take a small section of the image with simple structures, where

there are strong signal contents.

 Using the sample gray levels of the object and background, we can
construct an un-blurred image of the same size and characteristics as
the observed sub-image.
 Then we can extend the transfer function to the whole image.

 Let the observed sub-image be denoted by gs(x, y) and the processed

sub-image
ASRao 31
 Assuming negligible noise effect , we have

Gs (u, v)
H s (u, v)  
Fs (u, v)
 Based on the information of Position invariance, we can deduce
complete degradation function H(u,v) from the characteristics of
above function

 Suppose the radial plot of Hs(u,v) has the approximate shape of the
Gaussian curve , same information is used for H(u,v) in the larger scale.

ASRao 32
Estimation by Experimentation

 If the equipment , with which we obtained the degraded image is

available, it is possible in principle to obtain an accurate estimation of
the degradation function.
 Images similar to the degraded image can be acquired with various
system settings until they are as closely as possible to the image we
with to restore.
 The idea is to obtain the impulse response of the degradation by
imaging an impulse using the same systems settings.
 An impulse is simulated by bright dot of light as bright as possible to
reduce the effect of noise to negligible value..
 Then Fourier transform of an impulse is a constant

G(u, v)
H (u, v) 
A
Where A is constant describing strength of the impulse.
ASRao 34
Estimation by Mathematical Model

 Used when we know physical mechanism underlying the image

formation process that can be expressed mathematically.
 Mathematical model has been used very long.

 In some cases, the model can even take environmental conditions that
cause degradation.

 Example: Consider the physical characteristics of atmospheric

turbulence.
5
 k (u 2  v 2 ) 6
H (u , v)  e

K – is the constant depends on the nature of turbulence

Different Restoration Approaches

 Algebraic approaches
 Unconstrained Optimization
 Constrained Optimization

 Frequency Domain
 Inverse Filter
 Wiener (minimum mean square error) filter

ASRao 37
Algebraic Approach to Restoration
 Algebraic methods are very popular for image restoration because it
uses the concept of matrices and linear algebra instead of those
integrals.
 The objective of image restoration is to estimate the original image, f
from the degraded image, g using some knowledge or assumption
about H and η.
fˆ an estimate ,
 The objective of algebraic approach is to seek of the
original image, f from the degraded image, g such that a predefined
criterion function is minimized.
Unconstrained Method of Restoration
 This approach has no constraints .
 Here the approximate image of the original image is recovered by
minimizing the noise.
 The basic image degradation model

 The general image degradation model is given as:

g  Hf  n     (1)
 In the absence of any knowledge about the noise, h, the objective of
unconstrained restoration is to seek an estimate, of the original image, f
from the degraded image, g such that Hfˆ approximates g and the norm
of the noise term is minimized. That is fˆ is found such that
2
  g  Hfˆ
2
       ( 2) is minimum, where

  T , norm of 
2

g  Hfˆ
2

 g  Hfˆ  g  Hfˆ ,
T
norm of g  Hfˆ

 The minimization of equation (2) achieved by differentiating it with

respect to fˆ and equating the result to zero.

   2H g  Hfˆ  0       (3)
 
2
T
fˆ
solving equation (3) for fˆ
fˆ  H 1 g        ( 4)
 The frequency domain representation of Equation (4) is
G (u , v)
Fˆ (u , v)          (5)
H (u , v)

Where Fˆ (u , v), G (u , v) and H (u , v) are the

Fourier Transforms of f , g and h, respectively.

• In the presence of noise, equation (5) becomes

ˆ G (u , v) N (u , v)
F (u , v)           (6)
H (u , v) H (u , v)
where N (u , v) is the Fourier transform of the noise,  .

The disadvantages of the inverse filtering (unconstrained restoration) are:

(i) The complete knowledge about H and η is required and
(ii) The restoration process is very sensitive to noise. That is, the
restoration result is dominated by noise if H(u,v) is zero or small.
ASRao 42
Constrained Method of Restoration

 The basic image degradation model

 The general image degradation model is given as:

g  Hf  n     (1)
 The objective of constrained restoration is to seek an estimate, fˆ of the
original image, f from the degraded image, g such that the criterion
function
 2 2
J ( fˆ )  Qf    g  Hfˆ
2
         (2)
 
is minimum. The basic image degradation model
   T , norm of 
2

g  Hfˆ
2

 g  Hfˆ  g  Hfˆ ,
T
norm of ( g  Hfˆ )

 The minimization of Equation (2) is achieved by differentiating it w.r.t fˆ

and equating the result to zero.
 
 J ( fˆ )
 2QT Qfˆ  2H T ( g  Hfˆ )  0      (3)
fˆ
solving for fˆ


fˆ  H T H  QT Q 
1
H T g      ( 4)
where
 1 / 

 Equation (4) yields different solutions for different choices of Q.

ASRao 44
ASRao 45
ASRao 46
Inverse filtering

 Inverse filtering is a technique used in image restoration to remove the

effects of degradation, such as blurring or noise.

 Mathematically:
 Degraded Image = Original Image * PSF
 Restored Image = Degraded Image * (1/PSF)
where * represents the convolution operation.

ASRao 47
ASRao 48
ASRao 49
ASRao 50
ASRao 51
ASRao 52
ASRao 53
Least Mean Square Filtering (Wiener Filter)
 The Wiener filter is one approach which makes use of knowledge of the
statistical properties of the noise besides the degradation function.
 A Wiener filter is a linear filtering method used to restore images that
have been degraded by both noise and blurring. It attempts to remove
both noise as well as the blur.
 It's one of the first methods developed to reduce random noise in
images.
 Wiener filtering was one of the first methods developed to reduce
additive random noise in images.
 It works on the assumption that additive noise is a stationary random
process, independent of pixel location; the algorithm minimizes the
square error between the original and reconstructed images.
 The Weiner filter was proposed by Norbert Wiener in 1940. It was
published in 1949.
 Wiener filter is not an adaptive filter
ASRao 54
as it assumes input to be stationary.
 The Wiener filter works by:

Estimating the original image

Using a statistical model of the image and noise to estimate the original
image

Minimizing the mean square error

Finding the balance between removing noise and keeping image details by
minimizing the mean square error between the original and restored
images

To generate best estimate fˆ of f , we minimize

2 

 ˆ 2
e  E f  f 


where E is the Expected value.
ASRao 55
More details about the Wiener filter:
Assumptions
The Wiener filter assumes that the image and noise are random processes
that are uncorrelated and have known power spectra.
• f and n are uncorrelated
• f and/or n is zero mean
• gray levels in fˆ are a linear function of gray levels in f

Calculation
The Wiener filter is usually applied in the frequency domain. The Discrete
Fourier Transform (DFT) is used to obtain the image spectrum, which is then
multiplied by the Wiener filter. The inverse DFT is then used to obtain the
image estimate.
Performance
The Wiener filter is known for its performance in denoising white noise. It
performs better than other non-linear filters for images corrupted by white
noise. ASRao 56
The Fourier transform of the correlation is Power Spectral Density (PSD).
Let the PSD of the input image f(x,y) is Sf(u,v) and noise n(x,y) is Sn(u,v).

g ( x, y )  (h * f )( x, y )   ( x, y )
G (u , v)  H (u , v) F (u , v)  N (u , v)

Then the best estimate fˆ (u, v) is given by

ASRao 57
Multiply numerator and denominator by H(u,v)

ASRao (a)
58
 At frequencies (u,v) where the signal is much stronger than the noise, the
ISNR is 0, and the Wiener filter reduces to the inverse filter.
 At frequencies (u,v) where the signal is much weaker, the ISNR will be
large and the corresponding component G(u,v) will be attenuated
(note that the Wiener filter cannot reconstruct such components well)
 When there is no blurring, but only noise, we have:

 The Wiener filter requires us to know the ISNR at different frequencies.

 We usually do not have knowledge of this quantity, but we can provide
ASRao 59
some estimate
 For example, in most (photographic) images, the ISNR is low at lower
frequencies and high at higher frequencies.

 Why? Images typically have high energy at lower frequencies, and low
energy at higher frequencies. But noise is typically spread-spectrum.

 If the original image Sf(u,v) is unknown, equation (a) can be written as:

where K is specified constant. Here the noise is assume to be white

noise whose spectrum |N(u,v)|2 is a constant. Taking the inverse Fourier
Transform of F(u,v) gives the restored image in the Spatial domain.

ASRao 60
Limitations of Wiener Filter
Requires a priori knowledge: Wiener filters require accurate knowledge of
the input signal, noise signal, and true signal's power spectra. This can be
difficult to obtain, especially when the signal or noise is non-stationary or
non-Gaussian.

Linear: Wiener filters are linear, so they can't handle nonlinear phenomena
like harmonics, clipping, or intermodulation.

Frequency-domain smoothing: Wiener filters can introduce undesirable

effects like blurring, ringing, or overshooting due to frequency-domain
smoothing.
Can't reconstruct degraded components: Wiener filters can't reconstruct
frequency components that have been degraded by noise, and they can't
restore components for which H(u,v)=0.

Low accuracy: Wiener filters have high computation speed but low
accuracy, which can cause problems for quantitative computation.
ASRao 61
1. Assumes Stationary Noise: Wiener filter assumes that the noise is
stationary, meaning it doesn't change over time. If the noise is non-
stationary, the filter may not work well.
2. Requires Accurate Noise Model: Wiener filter needs an accurate model
of the noise to work effectively. If the noise model is incorrect, the filter
may not perform well.
3. Sensitive to Filter Parameters: Wiener filter's performance depends on
the choice of filter parameters. If these parameters are not chosen
correctly, the filter may not work well.
4. Not Effective for Non-Linear Distortions: Wiener filter is designed to
handle linear distortions, but it may not be effective for non-linear
distortions.
5. Computational Complexity: Wiener filter can be computationally
intensive, especially for large images or signals.
6. Assumes Gaussian Noise: Wiener filter assumes that the noise is
Gaussian, which may not always be the case. If the noise is non-Gaussian,
the filter may not work well.

ASRao 62
ASRao 63
Constrained Least Square Filtering
 Constrained least squares filtering (CLS) is an image processing technique
that reduces noise and degradation in an image
 Wiener filter is not optimal when no information on the power spectra is
available.
 For Wiener filters, the power spectra of the undegraded image (Sf) and
noise (Sn) must be known/assumed.
 CLSF does not require any assumption about original image, and only
makes use of mean and variance of the noise.
 The method to improve the Wiener filter is to impose a constraint on the
mean square error term. Hence CLSF is an advanced version of the
Wiener filter where we do not need the knowledge of Sf and Sn; and
some constraint is added to the least square filtering.
 The CLSF algorithm yields optimal result for each image to which it is
applied.

ASRao 64
 Degradation model: g(x,y) = f(x,y) * h(x,y) + η(x,y)

 Written in a matrix form: g = H f + η

 g, f, n matrices will have dimensions MN X 1

 H matrix will have MN X MN dimension

 Thus, the constrained least square filtering approach restores an image

using matrix calculations.

 There are two main problems in the matrix calculation of constrained

least square filtering. They are
1. Even for medium size images, the matrix dimensions will be very large
which makes the computation difficult.
2. The matrix H is highly sensitive to noise.

 The noise sensitivity can be reduced by performing the restoration based

on some measures of smoothness.
ASRao 65
 Objective is to find the minimum of a criterion function

 
M 1N 1 2
C     2 f ( x, y )
x 0 y 0
2
g  Hfˆ 
2
 Subject to the constraint, (b)

 The issue over here is to find the sensitivity of H to η i.e., blur to noise
and to solve this issue, we provide a constraint to the smoothing by using
the Laplacian operator that is second order derivative. Hence the
objective is to find the minimum of the criterion function given by C
which is equal to the Laplacian operator.
 The solution of this problem in frequency domain is
 H *
(u , v) 
ˆ
F (u , v)    G (u , v)
 H (u , v)   P (u , v) 
2 2

γ is a parameter that must be adjusted so that the constraint in above

equation is satisfied and P(u,v) is the Fourier transform of the function
ASRao 66
p(x,y)
where  0 1 0 
P(u,v) = Fourier transform of p(x,y) =  1 4  1
 0  1 0 

We recognize this function as the Laplacian operator.

 Spatial domain functions, must be properly padded with zeros prior to
computing their Fourier transforms.

 when γ =0, constrained least square filter reduces to Inverse filter, i.e.,

ˆ G (u , v)
F (u , v) 
H (u , v)

 It is easier to select the scalar value for γ than to approximate the SNR
which is seldom constant.

ASRao 67
 Image degraded by
Motion Blur + AWGN

 Low noise: Wiener and CLS generate equal results.

 High noise: CLS outperforms Wiener
ASRaoif γ is properly selected. 68
Constrained Least Squares Filter: Adjusting γ

Define a residual vector   g  Hfˆ

where fˆ and g are functions of γ.

Q(γ) is monotonically increasing function of γ, is defined as

Q ( )   T   
2

Adjust γ to satisfy this expression


2 2
 n  a (c)

where a is an accuracy factor.

If a=0,  2
 n
2
and the constrained in equation (b) will be satisfied.

ASRao 69
1. Specify an initial value of γ
2
2. Compute 

3. Stop if (c) is satisfied.


2 2
otherwise return to step 2 after increasing γ if  n a


2 2
or decreasing γ if  n a

Use the new value of γ to recompute

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

ASRao 70
C L S F Vs M M S E

 Wiener (MMSE) is optimal in the average sense while CLSF yields optimal
result to each image to which it is applied.

 When noise is low, both MMSE and CLSF generate equal results, but
when noise is high CLSF is advantageous if proper γ is selected. Because
it is easier to select scalar value of γ rather than to approximate the K
which is never constant.

Advantages
 Only mean and variance of the noise are required to be known to
implement this method.
 For each input image it produces ASRao
an optimal result. 71
ASRao 72
 Interactive restoration is a technique used in digital image processing to
remove noise, defects, or unwanted features from an image. This
process involves human interaction and feedback to achieve the desired
outcome.

 Why Interactive Restoration?

Improved accuracy: Human judgment and feedback can lead to more
accurate results, especially in cases where automated algorithms struggle.
Customization: Interactive restoration allows users to tailor the restoration
process to their specific needs and preferences.
Flexibility: This approach can be used for a wide range of image
restoration tasks, from simple noise removal to complex defect correction.

ASRao 73
Steps involved in Interactive Restoration
1. Image acquisition: The image to be restored is acquired and loaded into
the system.
2. Pre-processing: The image may undergo pre-processing steps such as
resizing, cropping, or converting to a suitable format.
3. Defect detection: The user identifies the defects or unwanted features
in the image.
4. Restoration: The user applies restoration techniques, such as filtering,
inpainting, or texture synthesis, to correct the defects.
5. Feedback and refinement: The user reviews the restored image and
provides feedback. The restoration process is refined based on this
feedback.
6. Final output: The final restored image
ASRao is generated. 74
Practical Examples
1. Noise removal: A user wants to remove noise from a digital image of a
landscape. They use an interactive restoration tool to select the noisy
areas and apply a noise reduction filter.
2. Scratch removal: A user wants to remove scratches from an old
photograph. They use an interactive restoration tool to select the
scratched areas and apply an inpainting algorithm to fill in the missing
information.
3. Text removal: A user wants to remove unwanted text from an image.
They use an interactive restoration tool to select the text and apply a
texture synthesis algorithm to replace it with a matching background
texture.
ASRao 75
Tools and Techniques
1. Image editing software: Adobe Photoshop, GIMP, etc.
2. Interactive restoration tools: Inpainting tools, texture synthesis tools,
etc.

Advantages and Limitations

Advantages:
- Improved accuracy and customization
- Flexibility in restoration techniques
- Human feedback and refinement
Limitations:
- Time-consuming and labor-intensive
- Requires user expertise and judgment
- May not be suitable for large-scale
ASRao image restoration tasks 76
 Interactive image restoration aims to generate restored images by
adjusting a controlling coefficient which determines the restoration level.

 Previous works are restricted in modulating image with a single

coefficient. In This approach the observer controls the restoration process
by tuning parameter and is able to obtain a final result that may be
adequate for a special purpose.

 Let η(x,y) denote a sinusoidal interference pattern of amplitude A and

frequency components (uo, vo)

(1)

Apply Fourier Transform on η(x,y)

ASRao (2)77
 The Fourier transform of 2D Sine function is a pair of impulses of strength
–A/2 and +A/2 located at coordinates (uo/2π. Vo/2π) and (-uo/2π. -Vo/2π)
of frequency plane.
 The transform has only imaginary component . If the degradation
considered is additive noise then the degraded image can be written as

G(u,v)=H(u,v) F(u,v) +N(u,v) (3)

 A display of the magnitude G(u,v) contains the magnitude of the sum of

F(u,v) and N(u,v).
 The construction of the filter H(u,v) requires considerable knowledge
about the image then f(x,y) can he obtained as
f(x,y) = F-1[H(u,v) G(u,v)] (4)

 The Bandpass filter is constructed interactively by observing of G(u,v) on a

display.
 If f(x,y) is known completely subtracting the pattern g(x,y) to obtain fˆ ( x, y )
ASRao 78
 The filtering procedure usually is only an approximation of the pattern
instead of subtracting f(x,y) from g(x,y) a weighted portion of f(x,y) can be
considered to obtain an estimate of
fˆ ( x, y )  g ( x, y )  w( x, y ) f ( x, y ) (5)

w(x,y)  weighted or modulation function

 One of the approach is to select w(x, y) so that the variance of

fˆ ( x, y ) is minimized over a specified neighborhood of every point (x, y).

 Consider a neighborhood of size (2x+1) by (2y+1) about a point (x,y).

 The local variance of fˆ ( x, y ) at coordinates (x,y) is

~ 2
   fˆ ( x  m, y  n)  fˆ ( x, y )
1 x y
 2 ( x, y ) 
(2 x  1)(2 y  1) m x n y  
(6)
where
~
fˆ ( x, y ) is the average value of fˆ ( x, y ) in the neighborhood
ASRao 79
 fˆ ( x, y ) 
1 x y

  fˆ ( x  m, y  n)
(2 x  1)(2 y  1) m x n  y

(7)
Sub equation 5 in equation 6
 g ( x  m, y  n)  
1 x y  
 ( x, y ) 
2
   w( x, y ) f ( x  m, y  n)  
(2 x  1)(2 y  1) m x n y  ~ 
~ 
 g ( x , y ) w( x , y ) f ( x , y ) 
(8)

 Differentiate w.r.t w(x,y) an equating to zero to minimize  2 ( x, y ) and

solving for w(x,y).

~ ~
g ( x, y ) f ( x, y )  g ( x, y ) f ( x, y )
w( x, y )  ~2
f ( x, y )  f ( x, y )
2

(9)
ASRao 80
 ~
 To obtain the restored image f ( x, y ) the weighted function w(x,y) is
computed . Generally w(x,y) is computed for one point in each non-
overlapping neighborhood and then used to process all the image points
in that neighborhood.

ASRao 81
ASRao 82