You are on page 1of 21

ENEE631 Digital Image Processing (Spring'04

)
Image Restoration
Spring ’04 Instructor: Min Wu

ECE Department, Univ. of Maryland, College Park
 www.ajconline.umd.edu (select ENEE631 S’04)
 minwu@eng.umd.edu
Based on ENEE631 Spring’04
Section 7
ENEE631 Digital Image Processing (Spring'04) Lec8 – Image Restoration [4]

From Matlab ImageToolbox
Documentation pp12-4
U
M
C
P

E
N
E
E
6
3
1

S
l
i
d
e
s

(
c
r
e
a
t
e
d

b
y

M
.
W
u

©

2
0
0
1
)

ENEE631 Digital Image Processing (Spring'04) Lec8 – Image Restoration [5]
Imperfectness in Image Capturing
 Blurring ~ linear spatial-invariant filter model w/ additive noise
 Impulse response h(n
1
, n
2
) & H(e
1
, e
2
)
– Point Spread Function (PSF) ~ positive I/O

– [No blur] h(n
1
, n
2
) = o(n
1
, n
2
)

– [Linear translational motion blur]
 local average along motion direction
– [Uniform out-of-focus blur]
 local average in a circular neighborhood
– Atomspheric turbulence blur, etc.
H
u(n
1
, n
2
) v(n
1
, n
2
)
N(n
1
, n
2
)
¦
¦
¦
¹
¦
¦
¦
´
¦
÷ =
s +
=
otherwise 0

tan and
x if
L
1
) , ; , ( y
x
2
2 2
|
|
L
y
L y x h
¦
¹
¦
´
¦
s +
=
otherwise 0
x if
R
1
) ; , (
2 2 2
2
R y
R y x h
t
U
M
C
P

E
N
E
E
6
3
1

S
l
i
d
e
s

(
c
r
e
a
t
e
d

b
y

M
.
W
u

©

2
0
0
1
)

ENEE631 Digital Image Processing (Spring'04) Lec8 – Image Restoration [6]
Fourier Transform of PSF for Common Distortions

From Bovik’s Handbook
Sec.3.5 Fig.2&3 UMCP ENEE631 Slides (created by M.Wu © 2001)
ENEE631 Digital Image Processing (Spring'04) Lec8 – Image Restoration [7]
Undo Linear Spatial-Invariant Distortion
 Assume noiseless and PSF of distortion h(n
1
, n
2
) is known






 Restoration by Deconvolution / Inverse-Filtering
– Often used for deblurring
– Want to find g(n
1
, n
2
) satisfies h(n
1
, n
2
) © g(n
1
, n
2
) = o(n
1
, n
2
)
 EE h(k
1
, k
2
) g(n
1
-k
1
, n
2
-k
2
) = o(n
1
, n
2
) for all n
1
, n
2


– Easy to solve in spectrum domain
 Convolution  Multiplication
 H(e
1
, e
2
) G(e
1
, e
2
) = 1
 Interpretation: choose G to compensate distortions from H
) , (
1
) , (
2 1
2 1
e e
e e
H
G =
H
u(n
1
, n
2
) v(n
1
, n
2
)
q(n
1
, n
2
)=0
G
u’(n
1
, n
2
) w(n
1
, n
2
)
U
M
C
P

E
N
E
E
6
3
1

S
l
i
d
e
s

(
c
r
e
a
t
e
d

b
y

M
.
W
u

©

2
0
0
1
)

ENEE631 Digital Image Processing (Spring'04) Lec8 – Image Restoration [8]
Problems With Inverse Filtering Under Noise
 Zeros in H(e
1
, e
2
)
– Interpretation: distortion by H removes all info. in those freq.
– Inverse filter tries to “compensate” by assigning infinite gains

– Amplifies noise:
 W(e
1
, e
2
) = H (e
1
, e
2
) U (e
1
, e
2
)
 U’ (e
1
, e
2
) = (W+N) / H ~ N/H if W=0

 Solutions?

U
M
C
P

E
N
E
E
6
3
1

S
l
i
d
e
s

(
c
r
e
a
t
e
d

b
y

M
.
W
u

©

2
0
0
1
)

ENEE631 Digital Image Processing (Spring'04) Lec8 – Image Restoration [9]
Examples of Inverse & Pseudo-inverse Filtering

From Jain Fig.8.10
U
M
C
P

E
N
E
E
6
3
1

S
l
i
d
e
s

(
c
r
e
a
t
e
d

b
y

M
.
W
u

©

2
0
0
1
)

ENEE631 Digital Image Processing (Spring'04) Lec8 – Image Restoration [10]
Problems With Inverse Filtering Under Noise
 Zeros in H(e
1
, e
2
)
– Interpretation: distortion by H removes all info. in those freq.
– Inverse filter tries to “compensate” by assigning infinite gains

– Amplifies noise:
 W(e
1
, e
2
) = H (e
1
, e
2
) U (e
1
, e
2
)
 U’ (e
1
, e
2
) = (W+N) / H ~ N/H if W=0

 Solutions ~ Pseudo-inverse Filtering
– Assign zero gain for G at spectrum nulls of H
– Interpretation: not bother to make impossible compensations

¦
¹
¦
´
¦
<
>
=
c e e
c e e
e e e e
| ) , ( H | if , 0
| ) , ( H | if ,
) , (
1
) , (
2 1
2 1
2 1 2 1
H G
U
M
C
P

E
N
E
E
6
3
1

S
l
i
d
e
s

(
c
r
e
a
t
e
d

b
y

M
.
W
u

©

2
0
0
1
)

ENEE631 Digital Image Processing (Spring'04) Lec8 – Image Restoration [11]
Handling Spectrum Nulls Via High-Freq Cut-off
 Limit the restoration
to lower frequency
components to avoid
amplifying noise at
spectrum nulls
U
M
C
P

E
N
E
E
6
3
1

S
l
i
d
e
s

(
c
r
e
a
t
e
d

b
y

M
.
W
u

©

2
0
0
4
)

Figure is from slides at Gonzalez/ Woods DIP book website (Chapter 5)
ENEE631 Digital Image Processing (Spring'04) Lec8 – Image Restoration [12]
Handling Noise in Deconvolution
 Inverse filtering is sensitive to noise
– Does not explicitly modeling and handling noise
 Try to balance between deblurring vs. noise suppression
– Minimize MSE between the original and restored
 e = E{ [ u(n
1
, n
2
) – u’(n
1
, n
2
) ]
2
} where u’(n
1
, n
2
) is a func. of {v(m
1
, m
2
) }

– Best estimate is conditional mean E[ u(n
1
, n
2
) | all v(m
1
, m
2
) ]
 usually difficult to solve for general restoration (need conditional
probability distribution, and estimation is nonlinear in general)
 Get the best linear estimate instead  Wiener filtering
– Consider the (desired) image and noise as random fields
– Produce a linear estimate from the observed image to minimize MSE
H
u(n
1
, n
2
) v(n
1
, n
2
)
q(n
1
, n
2
)
G
u’(n
1
, n
2
) w(n
1
, n
2
)
U
M
C
P

E
N
E
E
6
3
1

S
l
i
d
e
s

(
c
r
e
a
t
e
d

b
y

M
.
W
u

©

2
0
0
1
/
2
0
0
4
)

ENEE631 Digital Image Processing (Spring'04) Lec8 – Image Restoration [13]
Wiener Filtering
 Get the best linear estimate minimizing MSE
– Assume spatial-invariant filter u’(n
1
, n
2
) = g (n
1
, n
2
)© v(n
1
, n
2
)
– Assume wide-sense stationarity for original signal and noise
– Assume noise is zero-mean and uncorrelated with original signal
 Solutions
– Bring into orthogonal condition
E{ [ u(n
1
, n
2
) – u’(n
1
, n
2
) ] v(m
1
, m
2
) }=0
– Represent in correlation functions: R
uv
(k,l) = g(k,l) © R
vv
(k,l)
– Take DFT to get representation in power spectrum density
) , (
) , (
) , (
2 1
2 1
2 1
e e
e e
e e
vv
uv
S
S
G =
¦
¹
¦
´
¦
=
+ =
) , ( ) , ( ) , (
) , ( ) , ( ) , ( ) , ( : Note
2 1 2 1
*
2 1
2 1 2 1
2
2 1 2 1
e e e e e e
e e e e e e e e
qq
uu uv
uu vv
S H S
S S H S
) , ( ) , ( ) , (
) , ( ) , (

2 1 2 1
2
2 1
2 1 2 1
*
e e e e e e
e e e e
qq
S S H
S H
uu
uu
+
= U
M
C
P

E
N
E
E
6
3
1

S
l
i
d
e
s

(
c
r
e
a
t
e
d

b
y

M
.
W
u

©

2
0
0
1
)

ENEE631 Digital Image Processing (Spring'04) Lec8 – Image Restoration [14]
More on Wiener Filtering
 Balancing between two jobs
for deblurring noisy image
– HPF filter for de-blurring
(undo H distortion)
– LPF for suppressing noise
 Noiseless case ~ Sqq = 0 (inverse filter)
– Wiener filter becomes pseudo-inverse filter for S
qq
÷ 0




 No-blur case ~ H = 1 (Wiener Smoothing Filter)
– Zero-phase filter to attenuate noise according to SNR at each freq.
¦
¹
¦
´
¦
=
=
=
+
=
÷ ÷
0 | ) , ( H | if , 0
0 | ) , ( H | if ,
) , (
1
/ | |
) , (
2 1
2 1
2 1 0
2
*
0 2 1
e e
e e
e e e e
qq qq
qq
H
S S H
H
G
S
uv
S
uu
S H
S
H
G
*
wiener 2 1

1
) , (
qq
e e
+
=
1 ) , (
) , (
) , ( ) , (
) , (
) , (
2 1
2 1
2 1 2 1
2 1
1 wiener 2 1
+
=
+
=
=
e e
e e
e e e e
e e
e e
qq SNR
SNR
uu
uu
H
S
S
S S
S
G
U
M
C
P

E
N
E
E
6
3
1

S
l
i
d
e
s

(
c
r
e
a
t
e
d

b
y

M
.
W
u

©

2
0
0
1
)

ENEE631 Digital Image Processing (Spring'04) Lec8 – Image Restoration [15]
Comparisons

From Jain Fig.8.11
U
M
C
P

E
N
E
E
6
3
1

S
l
i
d
e
s

(
c
r
e
a
t
e
d

b
y

M
.
W
u

©

2
0
0
1
)

ENEE631 Digital Image Processing (Spring'04) Lec8 – Image Restoration [16]
Example: Wiener Filtering vs. Inverse Filtering
U
M
C
P

E
N
E
E
6
3
1

S
l
i
d
e
s

(
c
r
e
a
t
e
d

b
y

M
.
W
u

©

2
0
0
4
)

Figure is from slides at Gonzalez/ Woods DIP book website (Chapter 5)
ENEE631 Digital Image Processing (Spring'04) Lec8 – Image Restoration [17]
Example (2):

Wiener Filtering
vs.
Inverse Filtering
U
M
C
P

E
N
E
E
6
3
1

S
l
i
d
e
s

(
c
r
e
a
t
e
d

b
y

M
.
W
u

©

2
0
0
4
)

Figure is from slides at Gonzalez/ Woods
DIP book website (Chapter 5)
ENEE631 Digital Image Processing (Spring'04) Lec8 – Image Restoration [18]
To Explore Further 
U
M
C
P

E
N
E
E
6
3
1

S
l
i
d
e
s

(
c
r
e
a
t
e
d

b
y

M
.
W
u

©

2
0
0
4
)

ENEE631 Digital Image Processing (Spring'04) Lec8 – Image Restoration [19]
Review: Handling Noise in Deconvolution
 Inverse filtering is sensitive to noise
– Does not explicitly modeling and handling noise





 Try to balance between deblurring vs. noise suppression
– Consider the (desired) image and noise as random fields
– Minimize MSE between the original and restored
 e = E{ [ u(n
1
, n
2
) – u’(n
1
, n
2
) ]
2
} where u’(n
1
, n
2
) is a func. of {v(m
1
, m
2
) }

– Best estimate is conditional mean E[ u(n
1
, n
2
) | all v(m
1
, m
2
) ]
 Get the best linear estimate  Wiener filtering
– Produce a linear estimate from the observed image to minimize MSE
H
u(n
1
, n
2
) v(n
1
, n
2
)
q(n
1
, n
2
)
G
u’(n
1
, n
2
) w(n
1
, n
2
)
U
M
C
P

E
N
E
E
6
3
1

S
l
i
d
e
s

(
c
r
e
a
t
e
d

b
y

M
.
W
u

©

2
0
0
1
/
2
0
0
4
)

ENEE631 Digital Image Processing (Spring'04) Lec8 – Image Restoration [20]
Wiener Filter: From Theory to Practice
Recall: assumed p.s.d. of image & noise random fields and
freq. response of distortion filter are known

 Why make the assumptions?

 Are these reasonable assumptions?

 What do they imply in our implementation of Wiener filter?
U
M
C
P

E
N
E
E
6
3
1

S
l
i
d
e
s

(
c
r
e
a
t
e
d

b
y

M
.
W
u

©

2
0
0
4
)

ENEE631 Digital Image Processing (Spring'04) Lec8 – Image Restoration [21]
Wiener Filter: Issues to Be Addressed
 Wiener filter’s size
– Theoretically has infinite impulse response ~ require large-size DFTs
– Impose filter size constraint: find the best FIR that minimizes MSE

 Need to estimate power spectrum density of orig. signal
– Estimate p.s.d. of blurred image v and compensate variance due to noise
– Estimate p.s.d. from a set of representative images similar to the images
to be restored
– Or use statistical model for the orig. image and estimate parameters

– Constrained least square filter ~ see Jain’s Sec.8.8 & Gonzalez Sec.5.9
 Optimize smoothness in restored image (least-square of the rough
transitions)
 Constrain differences between blurred image and blurred version of
reconstructed image
 Estimate the restoration filter w/o the need of estimating p.s.d.

 Unknown distortion H ~ Blind Deconvolution
U
M
C
P

E
N
E
E
6
3
1

S
l
i
d
e
s

(
c
r
e
a
t
e
d

b
y

M
.
W
u

©

2
0
0
1
)

ENEE631 Digital Image Processing (Spring'04) Lec8 – Image Restoration [24]
Basic Ideas of Blind Deconvolution
 Three ways to estimate H: observation, experimentation, math. modeling
 Estimate H via spectrum’s zero patterns
– Two major classes of blur (motion blur and out-of-focus)
– H has nulls related to the type and the parameters of the blur
 Maximum-Likelihood blur estimation
– Each set of image model and blur parameters gives a “typical” blurred
output; Probability comes into picture because of the existence of noise
– Given the observation of blurred image, try to find the set of parameters that
is most likely to produce that blurred output
 Iteration ~ Expectation-Maximization approach (EM)
 Given estimated parameters, restore image via Wiener filtering
 Examine restored image and refine parameter estimation
 Get local optimums
 To explore more: Bovik’s Handbook Sec.3.5 (subsection-4, pp136)
“Blind Image Deconvolution” by Kundur et al, IEEE Sig. Proc. Magazine, vol.13, 1996
U
M
C
P

E
N
E
E
6
3
1

S
l
i
d
e
s

(
c
r
e
a
t
e
d

b
y

M
.
W
u

©

2
0
0
1
/
2
0
0
4
)

ENEE631 Digital Image Processing (Spring'04) Lec8 – Image Restoration [25]
Filtering Through Transform Domain Operation
E.g.1 Realize Wiener in DFT domain
E.g.2 Use zonal mask in transform domain
– Realize “ideal” LPF/BPF/HPF
– Computation complexity for transform could be high for large image
From Jain Fig.7.31&7.32
U
M
C
P

E
N
E
E
6
3
1

S
l
i
d
e
s

(
c
r
e
a
t
e
d

b
y

M
.
W
u

©

2
0
0
1
)

UMCP ENEE631 Slides (created by M.Wu © 2001)

From Matlab ImageToolbox Documentation pp12-4
ENEE631 Digital Image Processing (Spring'04) Lec8 – Image Restoration [4]

y. n2) & H(1.Imperfectness in Image Capturing u(n1. R)  R 0 otherwise  Lec8 – Image Restoration [5] .Wu © 2001) v(n1.  )     0 otherw ise  – [Uniform out-of-focus blur]  – Atomspheric turbulence blur. 2) – Point Spread Function (PSF) ~ positive I/O – [No blur] h(n1. n2 ) H N(n1. n2) – [Linear translational motion blur]  local average along motion direction local average in a circular neighborhood 1 if x 2  y 2  L 2 L   and x y   tan  h ( x. y. etc. L. n2) = (n1. ENEE631 Digital Image Processing (Spring'04)  1 2 2 2  2 if x  y  R h( x. n2 )   Blurring ~ linear spatial-invariant filter model w/ additive noise Impulse response h(n1. n2 ) UMCP ENEE631 Slides (created by M.

Wu © 2001) From Bovik’s Handbook Sec.5 Fig.2&3 Lec8 – Image Restoration [6] ENEE631 Digital Image Processing (Spring'04) .Fourier Transform of PSF for Common Distortions UMCP ENEE631 Slides (created by M.3.

n2) = (n1. n2) w(n1. n2) satisfies h(n1. 2) G(1. n2-k2) = (n1. n2)   h(k1. n2) u’(n1. n2) is known u(n1.  2 )  H (1 . n2 – Easy to solve in spectrum domain    Convolution  Multiplication H(1.Wu © 2001) Assume noiseless and PSF of distortion h(n1. 2) = 1 Interpretation: choose G to compensate distortions from H 1 G (1 .Undo Linear Spatial-Invariant Distortion  UMCP ENEE631 Slides (created by M.  2 ) ENEE631 Digital Image Processing (Spring'04) Lec8 – Image Restoration [7] . k2) g(n1 -k1. n2) for all n1. n2) v(n1. n2) H (n1. n2)=0  G Restoration by Deconvolution / Inverse-Filtering – Often used for deblurring – Want to find g(n1. n2)  g(n1.

2) U (1. in those freq. 2) = H (1.Problems With Inverse Filtering Under Noise UMCP ENEE631 Slides (created by M. 2) = (W+N) / H  N/H if W=0  Solutions? ENEE631 Digital Image Processing (Spring'04) Lec8 – Image Restoration [8] . – Inverse filter tries to “compensate” by assigning infinite gains – Amplifies noise:   W(1. 2) – Interpretation: distortion by H removes all info.Wu © 2001)  Zeros in H(1. 2) U’ (1.

8.10 Lec8 – Image Restoration [9] ENEE631 Digital Image Processing (Spring'04) .Wu © 2001) From Jain Fig.Examples of Inverse & Pseudo-inverse Filtering UMCP ENEE631 Slides (created by M.

2) U’ (1. 2) – Interpretation: distortion by H removes all info. 2 ) |   ENEE631 Digital Image Processing (Spring'04) Lec8 – Image Restoration [10] . 2 ) |   G (1 . if | H(1 . 2 )   H (1 . in those freq. – Inverse filter tries to “compensate” by assigning infinite gains – Amplifies noise:   W(1. 2) U (1. 2) = H (1. 2) = (W+N) / H  N/H if W=0  Solutions ~ Pseudo-inverse Filtering – Assign zero gain for G at spectrum nulls of H – Interpretation: not bother to make impossible compensations 1  .Wu © 2001) Zeros in H(1. if | H(1 .Problems With Inverse Filtering Under Noise  UMCP ENEE631 Slides (created by M. 2 ) 0.

Wu © 2004)  Limit the restoration to lower frequency components to avoid amplifying noise at spectrum nulls Figure is from slides at Gonzalez/ Woods DIP book website (Chapter 5) ENEE631 Digital Image Processing (Spring'04) Lec8 – Image Restoration [11] .Handling Spectrum Nulls Via High-Freq Cut-off UMCP ENEE631 Slides (created by M.

Wu © 2001/2004)  Inverse filtering is sensitive to noise – Does not explicitly modeling and handling noise  Try to balance between deblurring vs. n2) v(n1. n2) | all v(m1 .Handling Noise in Deconvolution UMCP ENEE631 Slides (created by M. n2) – u’(n1. m2) ]  usually difficult to solve for general restoration (need conditional probability distribution. noise suppression – Minimize MSE between the original and restored  e = E{ [ u(n1. m2) } – Best estimate is conditional mean E[ u(n1 . of {v(m1. n2) u’(n1. n2) H (n1. n2) w(n1. n2) is a func. and estimation is nonlinear in general)  Get the best linear estimate instead  Wiener filtering – Consider the (desired) image and noise as random fields – Produce a linear estimate from the observed image to minimize MSE u(n1. n2) ENEE631 Digital Image Processing (Spring'04) G Lec8 – Image Restoration [12] . n2) ] 2 } where u’(n1.

 2 ) S vv (1 .  2 ) Suu (1 .Wu © 2001) Get the best linear estimate minimizing MSE – Assume spatial-invariant filter u’(n1.  2 )  G (1 .  2 )  2 H (1 .l) – Take DFT to get representation in power spectrum density Suv (1 . n2) ] v(m1.  2 ) H * (1 .  2 )  H * (1 . n2) – Assume wide-sense stationarity for original signal and noise – Assume noise is zero-mean and uncorrelated with original signal  Solutions – Bring into orthogonal condition E{ [ u(n1.  2 ) Suu (1 . n2) – u’(n1.l) = g(k.  2 ) 2 Note : Svv (1 .Wiener Filtering  UMCP ENEE631 Slides (created by M. n2) v(n1.  2 )  H (1 .  2 )  S (1 . m2) }=0 – Represent in correlation functions: Ruv(k.  2 )   Suv (1 .  2 ) Suu (1 .  2 )  ENEE631 Digital Image Processing (Spring'04) Lec8 – Image Restoration [13] .  2 ) Suu (1 .l)  Rvv(k.  2 )  S (1 . n2) = g (n1.

2 ) | 0    H (1 .  2 ) wiener H 1  S uu (1 . if | H(1 . G (1 .  2 ) S SNR (1 .More on Wiener Filtering  UMCP ENEE631 Slides (created by M.  2 )  S (1 .  2 ) wiener  1 H  S H * S uu  Noiseless case ~ S = 0 (inverse filter) – Wiener filter becomes pseudo-inverse filter for S  0  H | H |2  S / Suv * G (1 . if | H(1 . 2 ) S  0 S  0 1  .  2 )  1 Lec8 – Image Restoration [14] ENEE631 Digital Image Processing (Spring'04) . 2 ) | 0   No-blur case ~ H = 1 (Wiener Smoothing Filter) – Zero-phase filter to attenuate noise according to SNR at each freq.Wu © 2001) Balancing between two jobs for deblurring noisy image – HPF filter for de-blurring (undo H distortion) – LPF for suppressing noise G (1 .  2 )  S uu (1 . 2 ) 0.  2 ) S SNR (1 .

8.Comparisons UMCP ENEE631 Slides (created by M.11 ENEE631 Digital Image Processing (Spring'04) Lec8 – Image Restoration [15] .Wu © 2001) From Jain Fig.

Example: Wiener Filtering vs. Inverse Filtering UMCP ENEE631 Slides (created by M.Wu © 2004) Figure is from slides at Gonzalez/ Woods DIP book website (Chapter 5) ENEE631 Digital Image Processing (Spring'04) Lec8 – Image Restoration [16] .

Example (2): UMCP ENEE631 Slides (created by M.Wu © 2004) Wiener Filtering vs. Inverse Filtering Figure is from slides at Gonzalez/ Woods DIP book website (Chapter 5) ENEE631 Digital Image Processing (Spring'04) Lec8 – Image Restoration [17] .

Wu © 2004) To Explore Further  ENEE631 Digital Image Processing (Spring'04) Lec8 – Image Restoration [18] .UMCP ENEE631 Slides (created by M.

Review: Handling Noise in Deconvolution UMCP ENEE631 Slides (created by M. n2) H (n1. m2) } – Best estimate is conditional mean E[ u(n1 .Wu © 2001/2004)  Inverse filtering is sensitive to noise – Does not explicitly modeling and handling noise u(n1. m2) ]  Get the best linear estimate  Wiener filtering – Produce a linear estimate from the observed image to minimize MSE ENEE631 Digital Image Processing (Spring'04) Lec8 – Image Restoration [19] . n2) v(n1. n2) ] 2 } where u’(n1. of {v(m1. n2) | all v(m1 . n2) is a func. noise suppression – Consider the (desired) image and noise as random fields – Minimize MSE between the original and restored  e = E{ [ u(n1. n2) – u’(n1. n2) u’(n1. n2) w(n1. n2) G  Try to balance between deblurring vs.

of image & noise random fields and freq.d.Wiener Filter: From Theory to Practice UMCP ENEE631 Slides (created by M.s.Wu © 2004) Recall: assumed p. response of distortion filter are known    Why make the assumptions? Are these reasonable assumptions? What do they imply in our implementation of Wiener filter? ENEE631 Digital Image Processing (Spring'04) Lec8 – Image Restoration [20] .

s.8 & Gonzalez Sec.d.9    Optimize smoothness in restored image (least-square of the rough transitions) Constrain differences between blurred image and blurred version of reconstructed image Estimate the restoration filter w/o the need of estimating p.s. of blurred image v and compensate variance due to noise – Estimate p.Wu © 2001) Wiener filter’s size – Theoretically has infinite impulse response ~ require large-size DFTs – Impose filter size constraint: find the best FIR that minimizes MSE  Need to estimate power spectrum density of orig. image and estimate parameters – Constrained least square filter ~ see Jain’s Sec.d.5. from a set of representative images similar to the images to be restored – Or use statistical model for the orig.8.s.  Unknown distortion H ~ Blind Deconvolution ENEE631 Digital Image Processing (Spring'04) Lec8 – Image Restoration [21] .Wiener Filter: Issues to Be Addressed  UMCP ENEE631 Slides (created by M. signal – Estimate p.d.

5 (subsection-4. Probability comes into picture because of the existence of noise – Given the observation of blurred image. math. pp136) “Blind Image Deconvolution” by Kundur et al. 1996 ENEE631 Digital Image Processing (Spring'04) Lec8 – Image Restoration [24] . experimentation. Magazine. vol.Basic Ideas of Blind Deconvolution  UMCP ENEE631 Slides (created by M.Wu © 2001/2004) Three ways to estimate H: observation. Proc. restore image via Wiener filtering Examine restored image and refine parameter estimation Get local optimums  To explore more: Bovik’s Handbook Sec. IEEE Sig.3. try to find the set of parameters that is most likely to produce that blurred output  Iteration ~ Expectation-Maximization approach (EM)    Given estimated parameters.13. modeling Estimate H via spectrum’s zero patterns – Two major classes of blur (motion blur and out-of-focus) – H has nulls related to the type and the parameters of the blur   Maximum-Likelihood blur estimation – Each set of image model and blur parameters gives a “typical” blurred output.

Filtering Through Transform Domain Operation UMCP ENEE631 Slides (created by M.32 Lec8 – Image Restoration [25] ENEE631 Digital Image Processing (Spring'04) .1 E.2 – – Realize Wiener in DFT domain Use zonal mask in transform domain Realize “ideal” LPF/BPF/HPF Computation complexity for transform could be high for large image From Jain Fig.7.g.31&7.Wu © 2001) E.g.