Image Restoration

Jayash Sharma
Department of Computer Science & Engineering
BMAS Engineering College, Agra
Email: jayash.sharma@gmail.com

What is Image Restoration:

Image restoration aim to improve an image in some
predefined sense.

What about image enhancement?

Image enhancement also improves an image by
applying filters.
Difference:

Image Enhancement --- Subjective process
Image Restoration --- Objective Process

Restoration tries to recover / restore degraded image by using a
prior knowledge of the degradation phenomenon.

Restoration techniques focuses on:

1. Modeling the degradation
2. Applying inverse process in order to recover the original image.
Model of the Image Degradation / Restoration Process

Degradation function along with some additive noise operates on
f(x, y) to produce degraded image g(x, y)

Given g(x, y), some knowledge about the degradation function
H and additive noise (x, y), objective of restoration is to obtain
estimate f(x, y) of the original image.

If H is linear, position invariant process then degraded image in
spatial domain is given by:
h(x, y) = Spatial representation of H
* indicates convolution
Since convolution in Spatial domain = multiplication in
Frequency Domain
We Assume that H is identity operator

We deal only with degradation due to Noise


Noise Models: Noise in digital image arises during

1. Image Acquisition
2. Transmission

During Image Acquisition

Environmental conditions (Light Levels)
Quality of sensing element

During Transmission

Interference during transmission
Spatial Properties of Noise:

1. With few exception we consider that noise is independent of spatial
coordinates.

2. We assume that noise is uncorrelated with respect to the image itself
(There is no correlation between image pixels and the values of noise
components)

Fourier Properties of Noise:

Refers to the frequency contents of noise in the Fourier sense.

If Fourier spectrum of noise is Constant, the noise is usually called
WHITE NOISE

Some Noise Probability Density Functions (PDFs):

Gaussian Noise
Rayleigh Noise
Erlang (Gamma) Noise
Exponential Noise
Uniform Noise
Impulse (Sal & Pepper Noise)
Periodic Noise

Spatial Noise Descriptor

Statistical behavior of the gray level values in the noise
component.

Can be considered as random variables

Characterized by Probability Density Functions (PDFs)



Gaussian / Normal Noise Model

1. Most frequently used.
2. PDF of Gaussian random variable z is given by:
z Gray level
Mean of average value of z
Standard Deviation of z

2
Variance of z
When z is defined by this equation then

About 70% of its values will be in the
range [( - ),( + )] and

About 95% of its values will be in the
range [( - 2),( + 2)]



Plot of function
Rayleigh Noise Model

PDF of Rayleigh Noise is given by:
z Gray level
Mean of average value of z

2
Variance of z

Basic shape of this density is skewed to
the right.

Quite useful for approximating skewed
histograms.
Plot of function
Erlang (Gamma) Noise Model

PDF of Erlang Noise is given by:
z Gray level
Mean of average value of z

2
Variance of z

Above equation is also called
Erlang Density

If denominator is Gamma function
then it is called Gamma density
Plot of function
a > 0
b = positive integer
Exponential Noise Model

PDF of Exponential Noise is given by:
z Gray level
Mean of average value of z

2
Variance of z
a > 0
Special case of Erlang Density
Where b=1
Plot of function
Uniform Noise Model

PDF of Uniform Noise is given by:
z Gray level
Mean of average value of z

2
Variance of z
Plot of function
Impulse (Salt & Pepper) Noise Model

PDF of Uniform Noise is given by:
z Gray level

If b > a then b light dot and a dark
dot

If either P
a
or P
b
= 0 Unipolar
Impulse Noise otherwise Bipolar
Impulse Noise.

If Neither probability is 0 and approximately equal then noise values will
resemble salt & pepper granules randomly distributed over the image.

Also referred as Shot and Spike Noise
Plot of function
This test pattern is well-suited for illustrating the noise models, because it is composed
of simple, constant areas that span the grey scale from black to white in only three
increments. This facilitates visual analysis of the characteristics of the various noise
components added to the image.
Example
Restoration using Spatial Filtering
We can use spatial filters of different kinds to remove different
kinds of noise.

Arithmetic Mean Filter
Let S
xy
represents the set of coordinates in a rectangular sub
image window of size m x n centred at (x, y).

This filter computes the average value of the corrupted image in
the area defined by S
xy.

e
=
xy
S t s
t s g
mn
y x f
) , (
) , (
1
) , (

1
/
9

1
/
9
1
/
9
1
/
9
1
/
9
1
/
9
1
/
9
1
/
9
1
/
9
Implemented as simple smoothing filter

Well Suited for Gaussian / Uniform Noise
Geometric Mean Filter

Each restored pixel is given by the product of the pixels in the
sub image window, raised to the power 1/mn.

Achieves smoothing comparable to Arithmetic Mean Filter but
tends to lose less image details in the process.

Well Suited for Gaussian / Uniform Noise

mn
S t s
xy
t s g y x f
1
) , (
) , ( ) , (

(
(

=
[
e
Harmonic Mean Filter

Works well for salt noise but fails for pepper noise.

Also does well for Gaussian Noise

e
=
xy
S t s
t s g
mn
y x f
) , (
) , (
1
) , (

Example:
Original
Image
Image
Corrupted
By Gaussian
Noise
After A 3*3
Geometric
Mean Filter
After A 3*3
Arithmetic
Mean Filter
Order Statistic Filter

Result is based on the ranking / ordering of the pixels contained
in the image area encompassed by the filter.

Median Filter
)} , ( { ) , (

) , (
t s g median y x f
xy
S t s e
=
Effective for both uni-polar and bipolar impulse noise.

Excellent at noise removal, without the smoothing effects that
can occur with other smoothing filters
Max Filter Good for Pepper Noise
Min Filter Good for Salt Noise

)} , ( { max ) , (

) , (
t s g y x f
xy
S t s e
=
)} , ( { min ) , (

) , (
t s g y x f
xy
S t s e
=
Mid Point Filter Good for Gaussian / Uniform Noise
(

+ =
e e
)} , ( { min )} , ( { max
2
1
) , (

) , ( ) , (
t s g t s g y x f
xy xy
S t s S t s
Image
Corrupted
By Salt And
Pepper Noise
Result of 1
Pass With A
3*3 Median
Filter
Result of 2
Passes With
A 3*3 Median
Filter
Result of 3
Passes With
A 3*3 Median
Filter
Example:
Image
Corrupted
By Pepper
Noise
Image
Corrupted
By Salt
Noise
Result Of
Filtering
Above
With A 3*3
Min Filter
Result Of
Filtering
Above
With A 3*3
Max Filter
Example:
Image
Corrupted
By Uniform
Noise
Image Further
Corrupted
By Salt and
Pepper Noise
Filtered By
5*5 Arithmetic
Mean Filter
Filtered By
5*5 Median
Filter
Filtered By
5*5 Geometric
Mean Filter
Filtered By
5*5 Alpha-Trimmed
Mean Filter
Example (Combined):
Periodic Noise
Typically arises due to electrical / Electro-
mechanical interference during image
acquisition.

Spatially dependent noise.

Can be reduced significantly via Frequency
Domain Filtering.

Parameters can be estimated by inspecting
the Frequency Spectrum of the image.

Periodic noise tend to produce frequency
spikes
Image corrupted by
Sinusoidal noise
Spectrum (Each pair of
conjugate impulses
corresponds to one sine wave)
Periodic Noise Reduction by Frequency Domain Filtering
Removing periodic noise form an image involves removing a
particular range of frequencies from that image

Bandreject Filter
Bandpass Filters
Notch Filter

Bandreject Filter
Removes / Attenuates a band of frequencies about the origin of the Fourier
Transform.

Ideal Bandreject Filter:

+ >
+ s s
<
=
2
) , ( 1
2
) , (
2
0
2
) , ( 1
) , (
0
0 0
0
W
D v u D if
W
D v u D
W
D if
W
D v u D if
v u H
D(u, v) = Distance of point from the origin
W = Width of the band
D0 = Radial Centre
Butterworth Bandreject Filter:
Gaussian Bandreject Filter:
Ideal Band
Reject Filter
Butterworth
Band Reject
Filter (of order 1)
Gaussian
Band Reject
Filter
Image corrupted by
sinusoidal noise
Fourier spectrum of
corrupted image
Butterworth band
reject filter
Filtered image
Example:
Inverse Filtering
An approach to restore an image.

Compute an estimate F( u, v) of the transform of the original image by:
Divisions are made between individual elements of the functions.
Inverse Filtering
Above Equation concludes that:

Even if we know degradation function, we can not recover the undegraded
image [Inverse Fourier Transform of F(u, v)] exactly because

N(u, v) is random function whose Fourier Transform is not known.

If degradation has ZERO or less value then N(u, v) / H(u, v) dominates the
estimated F(u, v).

No explicit provision for handling Noise.
Maximum Mean Square Error (Wiener) Filtering
Incorporates both degradation function and statistical characteristics of
noise into restoration process.

Considers images and noise as random process.

Find an estimate f of the uncorrupted image f such that mean square error
between them is minimized. Error measure is given by:


E{.} = Expected value of the argument

Assumptions:
image and noise are uncorrelated.
One or other has Zero mean
Gray levels in the estimate are a linear function of levels in the
degraded image.
Maximum Mean Square Error (Wiener) Filtering
Based on these conditions:


Singularity & Ill-condition?

Today After Lunch