A model of the image

degradation/restoration process

g(x,y)=f(x,y)*h(x,y)+h(x,y)
G(u,v)=F(u,v)H(u,v)+N(u,v)
 Noise models 雜訊模型
◼ Source of noise
◼ Image acquisition (digitization)
◼ Image transmission
◼ 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
◼ Ex. white noise (a constant Fourier spectrum)
 Noise probability density
functions
◼ Noises are taken as random variables
◼ Random variables
◼ Probability density function (PDF)
 Gaussian noise
◼ Math. tractability in spatial and frequency
domain
◼ Electronic circuit noise and sensor noise
1 − ( z −  ) 2 / 2 2
p( z ) = e
2 
mean variance

Note:
 p( z)dz = 1
−
Gaussian noise (PDF)

70% in [(−), (+)]

95% in [(−2), (+2)]
 Uniform noise
◼ Less practical, used for random number
generator
 1
 if a  z  b
p( z ) =  b − a
 0 otherwise
a+b
Mean:  =
2
(b − a ) 2

Variance:  =
2

12
Uniform PDF
 Impulse (salt-and-pepper)
nosie
◼ Quick transients, such as faulty switching during
imaging
 Pa for z = a

p ( z ) =  Pb for z = b
0 otherwise


If either Pa or Pb is zero, it is called unipolar.

Otherwise, it is called bipoloar.
•In practical, impulses are usually stronger than image
signals. Ex., a=0(black) and b=255(white) in 8-bit image.
Impulse (salt-and-pepper)
nosie PDF
 Test for noise behavior
◼ Test pattern

Its histogram:

0 255
 Periodic noise
◼ Arise from electrical or electromechanical
interference during image acquisition
◼ Spatial dependence
◼ Observed in the frequency domain
Sinusoidal noise:
Complex conjugate
pair in frequency
domain
 Estimation of noise parameters
◼ Periodic noise
◼ Observe the frequency spectrum
◼ Random noise with unknown PDFs
◼ Case 1: imaging system is available
◼ Capture images of “flat” environment
◼ Case 2: noisy images available
◼ Take a strip from constant area
◼ Draw the histogram and observe it
◼ Measure the mean and variance
Observe the histogram

Gaussian uniform
 Measure the mean and variance
◼ Histogram is an estimate of PDF

 =  zi p ( zi )
zi S
 2 =  ( zi −  ) 2 p ( zi )
zi S

Gaussian: , 
 Uniform: a, b
 Outline
◼ A model of the image degradation / restoration
process
◼ Noise models
◼ Restoration in the presence of noise only – spatial
filtering
◼ Periodic noise reduction by frequency domain
filtering
◼ Linear, position-invariant degradations
◼ Estimating the degradation function
◼ Inverse filtering
Additive noise only

g(x,y)=f(x,y)+h(x,y)
G(u,v)=F(u,v)+N(u,v)
 Spatial filters for de-noising
additive noise
◼ Skills similar to image enhancement
◼ Mean filters
◼ Order-statistics filters
◼ Adaptive filters
 Mean filters
◼ Arithmetic mean

ˆf ( x, y ) = 1
 g ( s, t )
mn ( s ,t )S xy
Window centered at (x,y)

◼ Geometric mean
1/ mn
ˆf ( x, y ) =   g ( s, t )
( s ,t )S xy 
 
original Noisy
Gaussian
=0
=20

Arith. Geometric
mean mean
 Mean filters (cont.)
◼ Harmonic mean filter
mn
fˆ ( x, y ) =
1

( s ,t )S xy g ( s, t )
◼ Contra-harmonic mean filter
 g ( s ,
( s ,t )S xy
t ) Q +1

fˆ ( x, y ) =
 g ( s , t ) Q

( s ,t )S xy Q=-1, harmonic

Q=0, airth. mean
Q=+, ?
Pepper Salt
Noise Noise
黑點 白點

Contra- Contra-
harmonic harmonic
Q=1.5 Q=-1.5
Wrong sign in contra-harmonic filtering

Q=-1.5 Q=1.5
 Order-statistics filters
◼ Based on the ordering(ranking) of pixels
◼ Suitable for unipolar or bipolar noise (salt and
pepper noise)
◼ Median filters
◼ Max/min filters
◼ Midpoint filters
◼ Alpha-trimmed mean filters
 Order-statistics filters
◼ Median filter
fˆ ( x, y ) = median{g ( s, t )}
( s ,t )S xy

◼ Max/min filters
fˆ ( x, y) = max {g ( s, t )}
( s ,t )S xy

fˆ ( x, y ) = min {g ( s, t )}
( s ,t )S xy
bipolar 3x3
Noise Median
Pa = 0.1 Filter
Pb = 0.1 Pass 1

3x3 3x3
Median Median
Filter Filter
Pass 2 Pass 3
Pepper Salt
noise noise

Max Min
filter filter
 Order-statistics filters (cont.)
◼ Midpoint filter
ˆf ( x, y ) = 1  max {g ( s, t )} + min {g ( s, t )}
2 ( s ,t )S xy ( s ,t )S xy 

◼ Alpha-trimmed mean filter

◼ Delete the d/2 lowest and d/2 highest gray-level
pixels
1
fˆ ( x, y ) =
mn − d
 g ( s, t )
( s ,t )S xy
r
Middle (mn-d) pixels
Uniform noise Left +
=0 Bipolar Noise
2=800 Pa = 0.1
Pb = 0.1

5x5 5x5
Arith. Mean Geometric
filter mean

5x5 5x5
Median Alpha-trim.
filter Filter
d=5
 Adaptive filters
◼ Adapted to the behavior based on the
statistical characteristics of the image inside
the filter region Sxy
◼ Improved performance v.s increased
complexity
◼ Example: Adaptive local noise reduction filter
 Adaptive local noise reduction
filter
◼ Simplest statistical measurement
◼ Mean and variance
◼ Known parameters on local region Sxy
◼ g(x,y): noisy image pixel value
◼ 2h: noise variance (assume known a prior)
◼ mL : local mean
◼ 2L : local variance
 Adaptive local noise reduction
filter (cont.)
◼ Analysis: we want to do
◼ If 2h is zero, return g(x,y)
◼ If 2L> 2h , return value close to g(x,y)
◼ If 2L= 2h , return the arithmetic mean mL
◼ Formula
ˆf ( x, y) = g ( x, y) −  h g ( x, y) − m 
2

L 2 L
Gaussian Arith.
noise
=0 mean
2=1000 7x7

Geometric
mean
7x7
adaptive
 Outline
◼ A model of the image degradation / restoration
process
◼ Noise models
◼ Restoration in the presence of noise only – spatial
filtering
◼ Periodic noise reduction by frequency domain
filtering
◼ Linear, position-invariant degradations
◼ Estimating the degradation function
◼ Inverse filtering
 Periodic noise reduction
◼ Pure sine wave
◼ Appear as a pair of impulse (conjugate) in the
frequency domain

f ( x, y) = A sin( u0 x + v0 y)
A u0 v0 u0 v0 
F (u, v) = − j  (u − , v − ) −  (u + ,v + )
2 2 2 2 2 
 Periodic noise reduction (cont.)
◼ Bandreject filters
◼ Bandpass filters
◼ Notch filters
◼ Optimum notch filtering
 Bandreject filters
* Reject an isotropic frequency

ideal Butterworth Gaussian

noisy spectrum
filtered

bandreject
 Bandpass filters
◼ Hbp(u,v)=1- Hbr(u,v)

−1 G (u , v) H bp (u , v)
 Notch filters
◼ Reject(or pass) frequencies in predefined
neighborhoods about a center frequency

ideal

Butterworth Gaussian
Horizontal
Scan lines

Notch
DFT pass

Notch Notch
pass reject
 Outline
◼ A model of the image degradation / restoration
process
◼ Noise models
◼ Restoration in the presence of noise only – spatial
filtering
◼ Periodic noise reduction by frequency domain
filtering
◼ Linear, position-invariant degradations
◼ Estimating the degradation function
◼ Inverse filtering
A model of the image
degradation /restoration process

g(x,y)=f(x,y)*h(x,y)+h(x,y)
G(u,v)=F(u,v)H(u,v)+N(u,v)

If linear, position-invariant system

 Linear, position-invariant
degradation
Properties of the degradation function H

◼ Linear system
◼ H[af1(x,y)+bf2(x,y)]=aH[f1(x,y)]+bH[f2(x,y)]
◼ Position(space)-invariant system
◼ H[f(x,y)]=g(x,y)
◼  H[f(x-a, y-b)]=g(x-a, y-b)
◼ c.f. 1-D signal
◼ LTI (linear time-invariant system)
 Linear, position-invariant
degradation model
◼ Linear system theory is ready
◼ Non-linear, position-dependent system
◼ May be general and more accurate
◼ Difficult to solve compuatationally
◼ Image restoration: find H(u,v) and apply
inverse process
◼ Image deconvolution
 Estimating the degradation
function
◼ Estimation by Image observation
◼ Estimation by experimentation
◼ Estimation by modeling
 Estimation by image
observation
◼ Take a window in the image
◼ Simple structure
◼ Strong signal content
◼ Estimate the original image in the window
known
Gs (u, v)
H s (u, v) =
Fˆs (u, v)
estimate
 Estimation by experimentation
◼ If the image acquisition system is ready
◼ Obtain the impulse response

impulse Impulse response

 Estimation by modeling (1)
− k ( u 2 + v 2 )5 / 6
◼ Ex. Atmospheric model H (u, v) = e

original k=0.0025

k=0.001 k=0.00025
 Estimation by modeling (2)
◼ Derive a mathematical model
◼ Ex. Motion of image
T
g ( x, y ) =  f ( x − x0 (t ), y − y0 (t ))dt
0

Fourier Planar motion

transform
T
G (u, v) = F (u, v)  e − j 2 [ ux0 ( t ) + vy0 ( t )]
dt
0
Estimation by modeling: example

original Apply motion model

 Inverse filtering
◼ With the estimated degradation function
H(u,v)
Unknown
G(u,v)=F(u,v)H(u,v)+N(u,v) noise

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

Estimate of Problem: 0 or small values

original image Sol: limit the frequency
around the origin
Full Cut
inverse Outside
filter 40%
for
k=0.0025

Cut Cut
Outside Outside
70% 85%