Beruflich Dokumente
Kultur Dokumente
i : pixel position
Goal
Remove noise while preserving the useful information.
Motivation
Preprocessing or joint processing for super resolution
imaging, image segmentation or quantitative analysis.
Construct an optimal estimation of the original signal based
on the observed image and the noise model.
Gaussian Noise
Noise is independent of signal with Gaussian pdf.
Poisson Noise
Noise variance is proportional to the signal.
Photon counting process is modeled with Poisson distribution.
When signal is large enough, the Poisson probability distribution
can be approximated as Gaussian pdf.
[Foi 2008]
The need for new noise modeling due to photon noise which is
signal-dependent.
z ( x) y ( x) p y ( x) g ( x)
Poisson signal-dependent
component
1
y ( x) p y ( x)
a
P y ( x) , g ( x)
a
N (0, b) (a 0, b 0)
1
1
E y ( x) p y ( x) var y ( x) p y ( x) y ( x)
a
a
a
1
1
1
1
E p y ( x) 0
E y ( x) p y ( x) y ( x) E p y ( x)
a
a
a
a
1
1
var
y
(
x
)
y ( x)
p
a
a
2 y( x) ay( x) b
Poisson signal-dependent
component
Wavelet transform
b=0.0034931
0.12
0.1
0.08
0.06
0.04
0.02
Observed image
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
0.1
0.08
0.06
0.04
0.02
Smoothed approximation
coefficients
Edge removed
detail coefficients
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
210000
10
20
30
800
40
50
60
70
80
0.08
Simulation noise
0.5
parameters
400
225
100
25
estimated a
-0.09
0.38
1.08
2.96
5.21
estimated b
430.94
255.88
101.71
47.37
estimated a
0.19
1.23
1.92
3.93
5.55
estimated b
464.45
246.08
101.16
8.70
estimated a
0.03
0.41
1.01
2.95
4.97
estimated b
404.34
240.91
101.93
25.00
0.32
Lena
Barbara
Camera man
zu
: equality in distribution
p(x | z ) pz ( z )dz
x T ( zC u ) 1 x
2
(2 ) N /2 zCu
1/2
[Portilla 2003]
pz ( z )dz
S (m) pW |S ( w | S
m 1
m).
[Crouse 1998]
Observation
Signal has high correlation in space domain
Noise components are independent
Solution
Spatial averaging
, M k M , N l N
0,
elsewhere
(m, n) ~ N (0, 2 )
Noise reduction
v(m, n)
1
u (m k , n l ) (m, n)
N w ( k ,l ) W
Nw
a(k , l )
1
Nw
0
1
1
4
1
4
1
4
1
4
0
1
(a) 2 x 2 window
1
9
1
9
1
9
1
9
1
9
1
9
1
9
1
9
1
9
(b) 3 x 3 window
1 0
1
8
1 0
0
1
8
1
2
1
8
1
0
1
8
Observation
Signal energy has much energy in low frequency.
Noise components have much energy in high
frequency range.
Solution
Filtering considering the noise and signal
spectrum
H * f1 , f 2
H f1 , f 2
1/[ Suu f1 , f 2 / S
f1 , f 2 ]
H f1 , f 2 : Image degradation
Suu f1 , f 2 / S
G f1 , f 2
when H
1 1/[ Suu
1
f1 , f 2 / S
f1 , f 2 ]
Domain filter
(Gaussian LPF)
Range filter
1
(m m0 ) 2 (n n0 ) 2
g[m, n] g[m0 , n0 ]
r
exp
exp
, [m, n] m0 , n0
2
2
h[m0 , n0 ; m, n] m0 , n0
2
d
r
, else
0
[m0 , n0 ] : center position for filtering
m0 , n0 : filtering window centered at [m0 , n0 ]
[Tomasi 1998]
p
pixel
intensity
q
domain
[Durand 02]
domain
pixel position
GB [ I ]p G || p q || I q
qS
domain
BF [ I ]p
1
Wp
pixel position
G || p q || G | I
qS
normalization
Observed
image
range
pixel
intensity
Gaussian LPF
Observed
image
domain
I q | I q
range
Bilateral filtering
Input image
Output image
*
*
(Domain filter)
[Durand 2002]
Input image
Output image
*
*
(Domain filter* Range filter)
Preserves edges
[Durand 2002]
Hard threshold
If an input is larger than threshold value, it is
maintained. Otherwise, the input is replaced to zero.
1 () I ( T )
w : wavelet coefficients,
T: threshold value
I : indicator function
Soft threshold
A method decreasing an input toward zero according to
a threshold value
2 () ( sgn()T ) I ( T )
The bias is larger than hard thresholds
Semisoft threshold
Complement the drawbacks of hard and soft thresholds
T2 ( T1 )
3 ( ) sgn( )
I (T1 T2 )
T2 T1
(T2 T1 0)
<Flow>
o vn
y w n'
v : original image
o : noisy image
n : Gaussian random variable
k ( yk ) arg max pw y ( wk yk )
w
Wavelet domain:
Observed image
Gaussian noise
Original image
w(y)
arg max py|w (y | w) pw (w).
w
Bayes rule
w(y)
arg max[
w
(y w)2
2 n2
w2
2 w2
w(y)
arg max[ pn (y w) pw (w)]
]
Logarithm
w2
w(y)
2
y
2
w n
[Mihcak 1999]
[Simoncelli 1999]
[Dabov 2007]
24c 7
E( f ( x))
var( f ( x))
1 3 8c 32c 2 52c 17
4
8
32 2
1/2
128
3/2
24c 7 16 2
E( f ( y ))
var( f ( y ))
1 3 8c 8 2 96c 2 52c 64 2 7
1
.
2
4
8
128
81/2
128 3/2
[Starck 1998]
[Zhang 2007]
1
2
( f (y ) 2y T f (y ) 2 2div{f ( y )})
N
1
2
( y 1T y ) 2
N
[Luisier 2011]
1
2
( f (y ) 2y T (f (y ) f (y )))
N
1
2
(2 2 div{f (y ) f (y )} y 1T y ) 2
N
approximation
Properties of wavelet
Clustering: If a wavelet coefficient is large/small, then adjacent
coefficients are very likely to also be large/small (horizontal
dependency)
Persistence: Large/small values of wavelet coefficients tend to
propagate across scales (vertical dependency)
(a) clustering
(b) persistence
[Crouse 1998]
m 1
(m) fW |S ( w | S m).
imr
, (i ) pSi |S ( i ) [m | S (i ) r ] : the conditional probability that Si
To search the HMM parameters that fit the given data best, an
expectation maximization (EM) algorithm is employed.
2
Denoise through Baysian estimator using the parameters pS1 (m), i ,m
2
i
E[ yik | w k , ] p( Sik m | w k , ) 2 ,m 2 wik
n i ,m
m
[1]
[Starck 2002]
<wavelet>
[Do 2005]
<contourlet>
[Do 2005]
zi k hk )2 ].
(z
n2i E[
i k
hk 2 )]
(E[ z
i k
]hk 2 )
2
2
zik hk )
z2ik hk2 .
E[vik
| w , ]
k
p( Sik
i2,m
k
m | w , ) 2
w
i
ni i2,m
k
1
s
K1K 2
^
K1 , K 2
i , j
i 1, j 1
Si , j
K1 , K2
Transform
Denoising
Inverse
Transform
(T 1 ( [T ( Si , j ( x)]))
w=y+n
<Conventional method>
x
WT
WGN
EM
Algorithm
Bayesian
Estimation
Using HMT
models
w=y + n
<PG-HMM>
x
noise
WT-1
EM
Algorithm
o x P ( x) g
CT
w CT (o)
y CT ( x)
n CT ( p ( x) g )
a, b estimation
for noise
variance
Bayesian
Estimation
Using HMT
models
Cycle spinning
Wiener
filtering
y
CT-1
Cell
Lena
Barbara
10
20
30
40
28.14
22.11
18.59
16.09
36.27
32.30
30.00
28.54
37.75
33.67
31.39
29.64
28.14
22.11
18.59
16.09
33.94
30.48
28.68
27.41
35.12
31.84
30.07
28.73
28.14
22.11
18.59
16.09
31.53
27.49
25.52
24.05
32.49
28.67
26.81
25.22
(a)
(b)
(c)
(d)
Results on cell image (a) original image (b) noisy image with n=20. Gaussian noise
(22.11 dB) (c) HMM based on wavelet (32.30 dB) (d) HMM based on contourlet (33.67 dB)
Lena
2.65
5.59
5.78
737.32
202
102
52
Noisy image
19.46
19.46
19.47
19.47
19.45
29.10
28.95
28.25
27.94
27.79
Gaussian contourlet
HMM
29.45
29.28
28.51
28.22
27.91
Gaussian contourlet
HMM with modified
Anscombe
29.09
28. 68
28.65
29.13
29.10
30.14
30.09
30.04
29.96
30.02
The proposed method shows the robust results regardless of noise rate
of Poisson and Gaussian.
Better performance than the conventional method with Anscombe.
Lena
Camera man
Cell
255
255/2
255/3
255/5
255/10
0.12
0.22
0.32
0.52
12 0
Noisy image
2.97
5.97
7.85
8.58
13.01
16.80
22.70
24.65
26.37
27.80
21.28
22.94
24.72
24.88
25.98
Noisy image
3.26
6.20
8.84
10.85
13.34
15.84
20.02
25.57
26.37
28.99
19.59
22.59
24.77
26.43
28.16
Noisy image
3.22
6.17
7.83
9.84
12.51
16.22
20.07
22.13
24.10
26.96
19.77
21.74
22.84
23.98
26.86
Original Image
(c)
(d)
(a)
(b)
(c)
(d)
Results on the magnified Lena image
(a) original image (b) noisy image with a=1/3, b=0.32 (c) BM3D (d) the proposed method
total
Noisy image
28.64
21.95
28.41
26.33
34.62
27.05
35.61
32.43
Bilateral filtering
34.35
26.45
36.81
32.54
32.69
25.23
39.97
32.63
35.58
29.53
36.94
34.02
(a)
(b)
(c)
(d)
(e)
(f)
1.
2.
3.
4.
5.
6.
7.
8.
9.
10. K. Dabov, A. Foi, V. Katkovnik, and K. Egiazarian, Image Denoising by Sparse 3-D Transform-Domain
Collaborative Filtering, IEEE Trans. on Image Processing, vol. 16, no. 8, pp. 2080-2095, Aug. 2007.
11. F. J. Anscombe, The Transformation of Poisson, Binomial and Negative-Binomial Data, Biometrika,
vol. 35, no. 3/4, pp. 246-254, Dec. 1948.
12. M. Fisz, The limiting distribution of a function of two independent random variables and its
statistical application, Colloq. Math., vol. 3, pp. 138-146, 1955.
13. D. L. Donoho, Nonlinear Wavelet Methods for Recovery of Signals, Densities, and Spectra from
Indirect and Noisy Data, in Proceedings of Symposia in Applied Mathematics, American
Mathematical Society, pp. 173-205, 1993.
14. P. Fryzlewicz and G. P. Nason, A Haar-Fisz Algorithm for Poisson Intensity Estimation, Journal of
Computational and Graphical Statistics, vol. 13, no. 3, pp. 621-638, 2004.
15. M. Jansen, Multiscale Poisson Data Smoothing, Journal of the Royal Statistical Society B, vol. 68, no.
1, pp. 2748, 2006.
16. B. Zhang, J. M. Fadili, and J.-L. Starck, J.-C. Olivo-Marin, Multiscale Variance-Stabilizing Transform for
Mixed-Poisson-Gaussian Processes and its Applications in Bioimaging, Proc. IEEE Int. Conf. Image Pro
cess., vol. 6, pp. 233-236, 2007.
17. B. Zhang, J. M. Fadili, and J.-L. Starck, Wavelets, Ridgelets, and Curvelets for Poisson Noise Removal,
IEEE Trans. on Image Processing, vol. 17, no. 7, pp. 10931108, 2008.
18. S. Lefkimmiatis, P. Maragos, and G. Papandreou, Bayesian inference on multiscale models for Poisson
intensity estimation: Applications to photon-limited image denoising, IEEE Trans. Image Processing,
vol. 18, no. 8, pp. 1724-1741, 2009.
19. F. Luisier, C. Vonesch, T. Blu, and M. Unser, Fast interscale wavelet denoising of Poisson-corrupted
images, Signal Processing, vol. 90, no. 2, pp. 415-427, 2010.
20. M. Makitalo and A. Foi, Optimal Inversion of the Anscombe Transformation in Low-count Poisson
Image Denoising, IEEE Trans. on Image Processing, vol. 20, no. 1, pp. 99-108, 2011.
21. D. L. Donoho, De-noising by soft-thresholding, IEEE Trans. Inform. Theory, vol. 41, pp. 613627, May
1995.
22. E. D. Kolaczyk, Nonparametric estimation of intensity maps using Haar wavelets and Poisson noise
characteristics, Astrophys. J., vol.534, pp. 490-505, 2000.
23. K. E. Timmermann and R. D. Nowak, Multiscale modeling and estimation of Poisson processes with
application to photon-limited imaging, IEEE Trans. Inf. Theory, vol. 45, no. 2, pp. 846-862, Apr. 1999.
24. E. D. Kolaczyk, Bayesian multi-scale models for Poisson processes, J. Amer. Stat. Assoc., vol. 94, no.
447, pp. 920-933, Sep. 1999.
25. H. Lu, Y. Kim, and J. M. M. Anderson, Improved Poisson intensity estimation: Denoising application
using Poisson data, IEEE Trans. Image Processing, vol. 13, no. 8, pp. 1128-1135, Aug. 2004.
26. S. Lefkimmiatis, P. Maragos, and G. Papandreou, Bayesian inference on multiscale models for Poisson
intensity estimation: Applications to photon-limited image denoising, IEEE Trans. Image Processing,
vol. 18, no. 8, pp. 1724-1741, Aug. 2009.
27. E. D. Kolaczyk, Wavelet shrinkage estimation of certain Poisson intensity signals using corrected
thresholds, Statistica Sinica, vol. 9, no. 1, pp. 119-135, Jan. 1999.
28. C. Charles and J. Rasson, Wavelet denoising of Poisson-distributed data and applications, Comput.
Stat. Data Anal., vol. 43, no. 2, pp. 139-148, 2003.
29. R. D. Nowak and R. G. Baraniuk, Wavelet-domain filtering for photon imaging systems, IEEE Trans.
Image Processing, vol. 8, no. 5, pp. 666-678, May 1999.
30. J.-L. Starck, F. Murtagh, and A. Bijaoui, Image Processing and Data Analysis, Cambridge University Pre
ss, 1998.
31. C. Stein, Estimation of the mean of a multivariate normal distribution, The Annals of Statistics, vol. 9,
pp. 11351151, 1981.
32. M. S. Crouse, R. D. Nowak, and R. G. Baraniuk, "Wavelet-based Statistical Signal Processing Using
Hidden Markov Models," IEEE Trans. on Signal Processing, vol. 64, no. 4, pp. 886-902, April 1998.
33. J. L. Starck, E. J. Cands, and D. L. Donoho, The Curvelet Transform for Image Denoising, IEEE Trans.
on Image Processing, vol. 11, no. 6, pp. 670-684, June 2002.
34. M. N. Do and M. Vetterli, The contourlet transform: An efficient directional multiresolution image
representation, IEEE Trans. on Image Processing, vol. 14, no. 12. pp. 2091-2106, Dec. 2005.
35. P. J. Burt and E. H. Adelson, The Laplacian pyramid as a compact image code, IEEE Trans. on
Communication, vol. 31, no. 4, pp. 532-540, April 1983.
36. D. D. Y. Po and M. N. Do, Directional multiscale modeling of image using the contourlet transform,
IEEE Trans. on Image Processing, vol. 15, no. 6, pp. 1610-1620, June 2006.
37. R. H. Bamberger and M. J. T. Smith, A Filter Bank for the Directional Decomposition of Images:
Theory and Design, IEEE Trans. on Signal Processing, vol. 40, no. 4, pp. 882-893, April 1992.
38. R. Eslami and H. Radha, The Contourlet Transform for Image De-noising Using Cycle Spinning, in
Proc. Asilomar Conference on Signal, Systems and Computer, pp. 1982-1986, November, 2003, Pacific
Grove, CA.
39. R. R. Coifman and D. L. Donoho, Translation Invariant Denoising, in Wavelets and Statistics, Springer
Lecture Notes in Statistics 103, New York, Springer-Verlag, pp. 125-150, 1994.
40. A. L. Cunha, J. Zhou and M. N. Do, The Nonsubsampled Contourlet Transform: Theory, Design, and
Applications, IEEE Trans. on Image Processing, vol. 15, no. 10, pp. 3089-3101, Oct. 2006.