IMAGE RESTORATION

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

Image restoration attempts to restore images that have been degraded

– Identify the degradation process and attempt to reverse it

– Similar to image enhancement, but more objective

Images are degraded due to noise. Noise may enter the image during image formation
process, the transmission process or a combination of both.

Images are degraded by noise, blur and distortion.

A model of the image degradation/restoration process

The process of image blur is usually complex. To simplify calculations the degradation is
often modelled as a linear function known as point spread function.

Where,
f(x,y) - input image
f^(x,y) - estimated original image
g(x,y) - degraded image
h(x,y) - degradation function
h(x,y) - additive noise term

The degraded image is given by

g(x,y)=f(x,y)*h(x,y)+h(x,y) – Spatial domain

Applying Fourier

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

The original image is restored by

 𝐺(𝑢, 𝑣) 𝑁(𝑢, 𝑣)
𝐹(𝑢, 𝑣) = −
𝐻(𝑢, 𝑣) 𝐻(𝑢, 𝑣)

NOISE MODELLING

Noise is a disturbance which fluctuates the pixel values. Since noise is a fluctuation in the pixel value,
it is considered a random variable. Thus noise categorization based on probability distribution is very
popular.

Sources of noise

– Image acquisition (digitization) - Imaging sensors can be affected by ambient conditions

– Image transmission - Interference can be added to an image during transmission

– Noise parameters, correlation with the image

Noise models
– Statistical behaviour of the Gray-level values of pixels

Spatial properties of noise

Frequency properties of noise

– Fourier spectrum
– Ex. white noise (a constant Fourier spectrum)

Gaussian noise

Mathematical tractability in spatial and frequency domains

Used frequently in practice

Gaussian noise
Electronic circuit noise and sensor noise, salt and pepper noise, shot noise
• Mathematical tractability in spatial and
Gaussian noise affects frequency
both the darkdomains
and the light areas of the image.

• Used
The Gaussian distribution frequently
is given by: in practice
• Electronic circuit noise and sensor noise
1 (z
2
) /2
2

p ( z) e
2

Intensity mean variance

Note: p ( z ) dz 1

Gaussian noise is also called additive noise and can be modelled by simple additive process.
This type of noise is caused due to sudden disturbance in the image signal. Mainly caused
by sensor and memory problems due to which pixels are assigned incorrect maximum values.
Gaussian noise PDF Gaussian noise PDF

70% in [( ), ( )]
95% in [( ), ( )]

Poisson Noise

Poisson distribution manifests itself in the form a random structures or textures in oamges and is very
common in x-ray images.

Poisson PDF

(𝑛𝑝)0 345
𝑃(𝑧) = 𝑒
𝑧!

n is the total number of pixels

p is the ratio of the noise pixels to total number of pixels

Exponential Noise:

This occurs due to illumination problems. The PDF is given by

𝑃(𝑧) = 6 𝑎𝑒 380 𝑖𝑓 𝑧 ≥ 0
0 𝑜𝑡ℎ𝑒𝑟𝑤𝑖𝑠𝑒

Rayleigh noise

Rayleigh noise
Mostly present in range images. Are used in remote sensing applications. The pixel values
here represent the distance between the object and the camera system.

2 2
( z a ) /b
(z a )e for z a
p ( z) b
0 for z a

–The mean and variance of this density are given

by
2 b(4 )
a b / 4 and
4
 Rayleigh noise
2 (z
2
a ) /b
(z a )e for z a
p ( z) b
0 for z a

–The mean
The and
mean and variance
variance of this
of this density aredensity
given by are given
by
2 b(4 )
a b / 4 and
4

a and b can be obtained through mean and variance

–a and b can
Gamma Erlang (Gamma) noise
Noisebe obtained through mean and
variance Erlang (Gamma) noise
Occurs due to illumination problem

b b 1
a z baz b 1
ea z for
e
az z 0
for z 0
p ( z) p ( z) (b 1)! ( b 1)!
0 0 for z 0 z
for 0

The mean–
– The mean and variance of thisof
andThe mean
variance and variance
of this density this density are given
are given by
density are given
by b / a and b
b 2

by b / a and a 2 2

2
a
– a and b can be obtained through mean and
a–
andab and
can be Uniform
variance
bobtained
can be obtained
through noise
mean and through mean and
variance

variance
Uniform noise
• Less practical, used for random number
Less practical, used for random number generator
generator
1
if a z b
p ( z) b a
0 otherwise
a b
Mean:
2
2
2 (b a)
Variance:
12

a and b are minimum and maximum value of the gray-scale value respectively
IMAGE BLUR

Caused due to

• improperly focussed lens

• relative motion between camera and scene
• atmospheric turbulence

Similar to noise, image blur can also be modelled.

GAUSS BLUR

PSF of Gauss Blur is

1 K L MN L
3
ℎ(𝑥, 𝑦) = 𝑒 OP L
𝜎√2𝜋

Gauss Blur occur due to long time atmospheric exposure.

OUT OF FOCUS BLUR

Caused due to defocussed optical system. It distributes a single point uniformity over a disk
surrounding the point.

ℎ(𝑥, 𝑦) = 𝑐 R1, S(𝑥 − 𝑐K )O − (𝑦 − 𝑐N )O ≤𝑟

0 𝑜𝑡ℎ𝑒𝑟𝑤𝑖𝑠𝑒

r is the radius, cx and cy is the centre out of focus PSF and c is a scaling factor.

Scaling factor is so chosen such that ∬ ℎ(𝑥, 𝑦)𝑑𝑥𝑑𝑦 = 1

MOTION BLUR

Due to movement of object or camera during light exposure, motion blur is produces. Different
types of motion blur

• translation
• rotation
• sudden change of scale
• combination of these

Length of motion is 𝐿 = 𝑣XYZ8[\]Y × 𝑡YK5_`aXY

The PSF is given by:

1 𝐿 𝑥
ℎ(𝑥, 𝑦, 𝐿, 𝜑) = 𝑐 R2 , c(𝑥) + (𝑦)
O O
≤ 𝑎𝑛𝑑 = 𝑡𝑎𝑛𝜑
2 𝑦
0 𝑜𝑡ℎ𝑒𝑟𝑤𝑖𝑠𝑒

F= angle between velocity and horizontal axis during exposure interval

Atmospheric Turbulence BLUR

This is a severe limitation in remote sensing.

Depends upon temperature, wind speed, exposure time (long time exposure). PSF is
given by

𝑥O + 𝑦O
ℎ(𝑥, 𝑦, 𝜎e ) = 𝑐𝑒𝑥𝑝 f− g
2𝜎e O

𝜎e 𝑑𝑒𝑡𝑒𝑟𝑚𝑖𝑛𝑒𝑠 𝑡ℎ𝑒 𝑎𝑚𝑜𝑢𝑛𝑡 𝑜𝑓 𝑠𝑝𝑟𝑒𝑎𝑑 𝑜𝑓 𝑏𝑙𝑢𝑟