Optics & Laser Technology 43 (2011) 926–933

Contents lists available at ScienceDirect

Optics & Laser Technology

journal homepage: www.elsevier.com/locate/optlastec

A piecewise local regularized Richardson–Lucy algorithm for remote sensing

image deconvolution
Wende Dong, Huajun Feng n, Zhihai Xu, Qi Li
State Key Laboratory of Modern Optical Instrumentation, Zhejiang University, Hangzhou 310027, China

a r t i c l e i n f o a b s t r a c t

Article history: As it is not possible to obtain an accurate point spread function (PSF) in remote sensing imaging, classic
Received 31 May 2010 deconvolution methods such as Wiener filtering often introduce strong noise and ringing artifacts, which
Received in revised form contaminate the restored images. In this paper, we modify the standard Richarson–Lucy (RL) algorithm
12 December 2010
with a piecewise local regularization term and combine it with residual deconvolution method.
Accepted 14 December 2010
Available online 30 December 2010
Experimental results show that it is effective in suppressing negative effects, and images with rich
details and sharp edges are obtained.
Keywords: Crown Copyright & 2010 Published by Elsevier Ltd. All rights reserved.
Image deconvolution
Richardson–Lucy algorithm
Regularization

1. Introduction estimated following the probabilistic framework. For example, RL

algorithm [3,4] assumes that the conditional likelihood probability
In remote sensing imaging, the obtained images are often of g given o follows a Poisson distribution and the author of [5]
blurred due to the relative movements and atmospheric distur- proposes a method based on Gaussian distribution. Fourier domain
bances. In addition devices such as CCDs also add noise. The image deconvolution methods are also commonly used, a typical example
deterioration process is described by is Wiener filtering [6]:

g ¼ h  o þ n ¼ Ho þ n ð1Þ ~  ðuÞ
H
~
OðuÞ ¼ ~
GðuÞ ð3Þ
2
where g is the observed image, h denotes the PSF, o is the latent
~
9HðuÞ9 þ S~ Z ðuÞ=S~ f ðuÞ
image,  is the convolution operator, H is the convolution matrix where H ~  ðuÞ is the conjugate of HðuÞ,
~ S~ Z ðuÞ and S~ f ðuÞ are the
and n stands for the additive noise. In frequency domain, Eq. (1) is respective power spectrums of noise and latent image. Other
transformed into the following expression: methods such as maximum entropy [7] and maximum expectation
~ ~ ~ ~ [8] are also included, see [9] for a summary.
GðuÞ ¼ HðuÞ OðuÞ þ NðuÞ ð2Þ
In blind deconvolution, the PSF is unknown. Multiplicative iter-
where G,~ H,
~ O~ and N~ denote the discrete Fourier Transforms of g, h, o ative algorithms are often used to estimate the latent image and PSF
and n, respectively. progressively [10–12]. In recent years, some progress has been
Image deconvolution methods are used to estimate the latent made in this field. For example, the method in [13] assumes that the
image from the degraded image. They can be divided into two gradient of the latent image follows a heavy tailed distribution and
categories, non-blind and blind deconvolution. In non-blind decon- estimates the PSF in scale space. In [14], the authors introduce a
volution, the PSF is known and o can be restored through an error probabilistic model, which combines both the PSF and image esti-
2
minimization process o ¼ arg mino :Hog: , where o* denotes the mation. In [15], the relationship between boundary transparency
estimated image. This problem is ill-conditioned and researchers and motion blur is studied. All these methods can be used to deblur
have developed various regularization methods to reach a stable remote sensing images.
solution. For example, in Tikhonov regularization [1], the regular-
2 2 2
ization term l :o: is suggested and o ¼ arg mino ð:Hog: þ 1.1. Problem formulation
2 2
l :o: Þ, methods such as conjugate gradient or Landweber [2]
can be used to iteratively update the estimated image. O can also be In image deconvolution, even if the PSF is known, some negative
effects will degrade the estimated images, the most typical are
n
Corresponding author. Tel./fax: + 86 571 87951182. amplified noise and ringing artifacts. This is because in frequency
E-mail address: [email protected] (H. Feng). ~
domain, the PSF is low pass and we can assume that HðuÞ  0 above

0030-3992/$ - see front matter Crown Copyright & 2010 Published by Elsevier Ltd. All rights reserved.
doi:10.1016/j.optlastec.2010.12.012
 W. Dong et al. / Optics & Laser Technology 43 (2011) 926–933 927

a cut-off frequency. Taking Wiener filter as an example show that our approach is very effective in restoring remote sen-
sing image.
~  ðuÞ
H ~  ðuÞ
H
~
OðuÞ ¼ ~
GðuÞ ¼ ~
HðuÞ ~
OðuÞ
2 2
~
9HðuÞ9 þ S~ Z ðuÞ=S~ f ðuÞ ~
9HðuÞ9 þ S~ Z ðuÞ=S~ f ðuÞ

~  ðuÞ
H 2. Our approach
þ ~
NðuÞ ð4Þ
2
~
9HðuÞ9 þ S~ Z ðuÞ=S~ f ðuÞ
In the field of image deconvoluiton, many approaches use the
Bayesian probabilistic framework:
where
o ¼ arg maxo Pðo9gÞpPðg9oÞPðoÞ ð5Þ
~  ðuÞ
H ~  ðuÞ
H
~  ðuÞOðuÞ
H ~ and ~
NðuÞ Take RL algorithm as an example, the Poisson process is applied
2 2
~
9HðuÞ9 þ S~ Z ðuÞ=S~ f ðuÞ ~
9HðuÞ9 þ S~ Z ðuÞ=S~ f ðuÞ to model the observed image. Suppose that every pixel in the image
are the respective frequency-domain expressions of estimated is independent, we obtain the following equation:
image and noise. Since NðuÞ ~ usually falls into the high-frequency ( )
Y ðh  oÞðxÞgðxÞ exp½h  oðxÞ
region where HðuÞ ~  0, the noise power will be amplified. Ringing Pðg9oÞ ¼ ð6Þ
gðxÞ!
artifacts are due to the well-known Gibbs phenomena. When Fourier x

series is used to approximate a signal, there will be high-frequency

where x is the pixel index. After logarithmic operation, Eq. (6)
fluctuations around the discontinuous points, these components will
becomes
also be amplified like the noise and alternating light and dark strips X
emerge around the image edges. These artifacts damage the details of J1 ðoÞ ¼ ln Pðg9oÞ ¼ fgðxÞln½ðh  oÞðxÞðh  oÞðxÞln½gðxÞ!g ð7Þ
the restored image. x

Many algorithms have been proposed to solve these problems. In order to reach a Maximum Likelihood solution of o, we set the
For example, in Total Variation (TV) regularization method [16–18], P
derivative of J1(o) to be zero, with x hðxÞ ¼ 1 we get
researchers introduce a gradient based regularization term Jt ðoÞ ¼  
qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi g
P 2
x :ro: þ b (where b is a tunable parameter) and estimate the  h ¼ 1 ð8Þ
 qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi oh
2 P 2
latent image by o ¼ arg mino 12 :Hog: þ l x :ro: þ b . In Richardson [3] and Lucy [4] suggested a multiplicative iterative
wavelet domain, natural image can be economically represented method to solve Eq. (8), which is given by:
by a few large coefficients while noise distributes over a large number   
g
of small ones, so the wavelet coefficients of the Fourier-based decon- on þ 1 ¼ h  on ð9Þ
on  h
volution result can be shrunk to get a noise-free image [19–21]. In
[22], a sparse prior is used to concentrate the derivatives only on a few where h*(i,j) ¼h(  j,  i) and n are the numbers of iterations. Since
pixels and get good results. For more methods, please refer to [23–25]. the inverse process is ill-conditioned, negative effects will emerge
However, all the above methods are effective only when the size as n increases, especially when n-+N, there is only noise left.
of the PSF is small (less than 15 pixels) and the pattern of its energy
distribution is simple. If the PSF is large or complex, these methods
cannot locate the discontinuities in the blurred image and lose their 2.1. Our regularized RL algorithm
efficacy. For example, TV regularization method will destroy lots of
details in the restored image. In actual remote sensing imaging, we In order to give out our regularization term, we first define a
can only get an approximation of the PSF and thus the noise and square spatial support S. We also define a series of displacement
ringing artifacts will be more serious, it will be difficult to obtain matrices Hki ði ¼ 1,2,   Þ that o Hki shifts the whole image o by 9ki9
high quality images with these methods. pixels along the direction of ki, where ki denotes the vector from the
center pixel to other pixels in S.
Fig. 1(a) shows a spatial support whose size is 3  3 and how
1.2. Overview of our method each ki is defined. Fig. 1(c) and (d) are two examples that show the
behavior of o Hki . The margin caused by shifting the image is filled
We introduce a new regularized RL algorithm to solve the above with zeros.
problems. First we design a local regularization term and then Our regularization term is defined as follows:
modify it with a piecewise function. This new term helps the RL
algorithm to accurately locate the discontinuities during the decon- Jr ðoÞ ¼ oT Mo ð10Þ
volution process and thus significantly suppresses the noise and X
ringing artifacts. We also combine this new algorithm with residual M¼ ðIHki Þ ð11Þ
deconvolution [26] to strengthen the edges. Experimental results i

k1 k2 k3
k4 k5

k6 k7 k8

Fig. 1. (a) definition of ki in a 3  3 spatial support, (b) a supposed simple image o, (c) the result of o Hk1 and (d) the result of o Hk2 .
928 W. Dong et al. / Optics & Laser Technology 43 (2011) 926–933

where I is an identity matrix, we add this term to J1(o) and obtain modified to put gentle penalties on the regions rich of details.
the following expression: The following piecewise function is capable to do this:
(
l ðx=255Þscale if 0 rt1 rðx=255Þ r t2
JðoÞ ¼ J1 ðoÞ þ Jr ðoÞ ð12Þ f ðxÞ ¼ ð22Þ
2
0 else
where l is the regularization coefficient. Suppose o* is the Max-
imum Likelihood solution of Eq. (12), we obtain where scaleA[  1,1], t1 and t2 are two thresholds.
  Using f(Mon) to replace Mon in Eq. (16), we obtain a new iterative
l expression:
o ¼ arg maxo J1 ðoÞ þ Jr ðoÞ ð13Þ
2   
on g
Take the derivative of Eq. (12), with (qJ(o)/qo)9O ¼ O* ¼ 0 and the on þ 1 ¼ h  ð23Þ
1 þ lf ðMon Þ on  h
One-Step-Late method proposed in [16,27], we get the following
iterative equation: We call this new method piecewise local regularized RL algo-
   rithm (PLRL).
on g
on þ 1 ¼ h  ð14Þ
1 þ lD½Jr ðon Þ=2 on  h
2.2. Residual deconvolution
where D[Jr(on)] ¼(q[Jr(on)]/qon), from the matrix derivation rules we
get
When we use PLRL to restore a blurred image, the edges will be
D½Jr ðon Þ ¼ ðM T þMÞon ð15Þ weakened to some degree, so they need to be strengthened. Here
we use the residual deconvolution method proposed in [26]. Denote
because M is symmetric, Eq. (14) becomes
   the result of PLRL as op, the residual blurred image is given by
on g
on þ 1 ¼ h  ð16Þ Dg ¼ gop  h ð24Þ
1 þ lMon on  h
In [26], standard RL algorithm is used to estimate the residual
where Mon can be efficiently calculated by:
X image. Here we substitute standard RL with PLRL. The residual
Mon ¼ ðon on Hki Þ ð17Þ blurred image Dg should be offset first by adding the constant 255
i to enforce the non-negativity, and the result must be offset back by
The regularization term Jr(o) can also be interpreted from the subtracting 255, the iteration is
perspective of Gaussian Markov Random Fields (GMRFs) [9,28,29].   
ðDon þ 255Þ ðDg þ255Þ
Let us have a two-dimensional lattice of sites C (the sites in C are Don þ 1 ¼ h  255
1 þ lf ½MðDon þ 255Þ ðDon þ255Þ  h
ordered by 1,2,. . .L) endowed with a given non-directed graph
ð25Þ
structure. If two sites i,jAC are neighbors, we note their neighbor-
hood relationship by i j. The definitions of non-directed graph and the final result is
neighbor are included in [30].
o ¼ op þ Don þ 1 ð26Þ
Let the mean of each pixel in o be zero. Then a GMRF on the joint
distribution of o is given by
PðoÞp expðoT AoÞ ð18Þ 3. Simulation
where the only non-null elements of A are
( Since our approach is non-blind, the PSF must be acquired
aii
ð19Þ through other ways. In order to imitate the actual situation, we
aij if i j
build an optical system.
Then the Hammersley–Clifford theorem [29,30] ensures that
the local-conditional distribution at any site i writes
2 0 13
X
4 @ 2
Pðoi 9foj gj i Þp exp  aii oi þ aij oi oj A5 ð20Þ
j i

In GMRF, the neighborhood structure is decided by the user, so

we can assume that the neighborhood of each site is the same as
S, i.e., for a given site centering at a square spatial support, all the
other sites in the support are its neighbors. If we fix aii ¼ 1
ði ¼ 1,2. . .Þ and aij ¼ 1=N with i j (where N denotes the number
of sites in the neighborhood), take the derivative of ln½PðoÞ, since A
is symmetric, we obtain:
Dfln½PðoÞgpðAT þAÞo ¼ 2Ao ð21Þ
P i
So Dfln½PðoÞgpNAo ¼ i ðooHg Þ, where g in GMRF is defined
the same as k. We can see that ln½PðoÞ is essentially equivalent to
the regularization term Jr(o).
Contrasting to other regularization term, such as Jt ðoÞ ¼
qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
P 2 P
x :ro: þ b in TV method, the new term Jr ðoÞ ¼ x oT Mo uses
more neighbor pixels in a local space S, which helps the RL
algorithm to locate the image discontinuities more accurately.
However, simulations show that this term puts a heavy penalty on
the entire image and damages the image details. It must be Fig. 2. Simulation system.
 W. Dong et al. / Optics & Laser Technology 43 (2011) 926–933 929

3.1. Obtaining the PSF piezoelectric controller drives the compensation mirror to keep the
image plane stable on the imaging CCD.
There are lots of ways to obtain the PSF. For example, in remote
sensing imaging, an accelerometer or a gyroscope can be used. In
3.2. Parameters setting
[13], the PSF is estimated with the natural image gradient prior. In
[26], blurred/noisy image pairs are used. The method in [31] adopts
Our method consists of the following steps:
the knife-edge method. In [32], a high-speed camera is attached to a
primary camera to detect the motion trajectory, the PSF is
estimated with the displacement vectors. These methods can 1. Use PLRL to obtain a preliminary result op;
provide an approximation of the PSF. 2. Get the residual blurred image Dg from g hop and deconvo-
We adopt the method suggested in [26] and [32] to estimate the lute Dg with PLRL to reach a residual image Do, the final result is
PSF. An optical system is built to simulate the image deterioration op + Do.
process. Fig. 2 is the system sketch. The system contains a high-
speed CCD module to realize the algorithm in [32]. The two- From Section 2, we know that there are several parameters need
dimensional vibration platform is used to simulate the relative to be set: the regularization coefficient l, the size of the spatial
movements between the camera and the target. The collimator is support S, two thresholds t1, t2 and the scale. Since in a blurred
used to imitate the parallel light path in remote sensing imaging. image, the latent clear edges are expanded by the scale of the PSF, S
The motion trajectory is detected by the high-speed CCD and the should be larger or at least of the same size as the PSF to ensure the

Fig. 3. Image restored with our method: (a) blurred image and the PSF (13  13), the PSF is estimated with the method in [32], (b) result of PLRL, (c) residual image restored by
residual deconvolution (it has been enhanced), (d) final result, (e)–(g) zoom of the restored images in the blue square using standard RL, PLRL and PLRL combining with residual
deconvolution. (For interpretation of the references to color in this figure legend, the reader is referred to the web version of this article.)
930 W. Dong et al. / Optics & Laser Technology 43 (2011) 926–933

discontinuities can be accurately located. But a too large S will Fig. 3 exhibits the result of each step consisting of our method.
remove too much image details. So we fix the size of S equals to the From Fig. 3(e) and (f) we can see that contrasting to standard RL,
PSF. For example, if the size of the PSF is 5  5, the size of S is set to PLRL significantly suppresses the noise and ringing artifacts. From
be 5  5. Fig. 3(c), (f) and (g) we see that the residual deconvolution
In step 1, the function of PLRL is to preserve as many as details enhances the image edges restored by PLRL.
while in step 2 it is used to strengthen the edges, so the remaining In Figs. 4 and 5, various algorithms are compared. Since the
parameters must be set separately. We empirically set t1 ¼1,t2 ¼ PSFs are not accurate, the results of Wiener filtering and standard RL
max½ðM T þ MÞon , scale¼0.1in each iteration of step 1 and t1 ¼0, are contaminated by noise and ringing artifacts. Although Levin’s
t2 ¼ 1=2max½ðM T þ MÞon , scale¼0 in each iteration of step 2. l is fixed method and TV regularization suppress these effects, image details are
to 0.01 in both steps. damaged. Our method outperforms these algorithms in not only
From experiments we learn that if the restored image is over suppressing negative artifacts but also preserving image details.
smoothed, t1 should be tuned larger and scale should be tuned
smaller. But if there is much noise, they should be tuned in the 4.2. Image quality assessment
opposite direction. l is usually tuned between 0.001 and 0.01. The
number of iterations is 40–60. In this section, we use SNR and another popular method:
Structural Similarity Index (SSIM) [33] to objectively evaluate
the restored images in Fig. 5. The results are given in Table 1.
4. Experimental results and evaluation SSIM is a widely used full-reference image quality assessment
method under the assumption that human visual perception is highly
In this section, we first use the system and method proposed in adapted for extracting structural information. SSIM combines three
Section 3 to obtain two blurred images and their PSFs. Then we components: local luminance, local contrast and structure to assess
deconvolute them with different algorithms. The results are evaluated an image. Eq. (27) is the definition of SSIM.
with two image assessment methods.
SSIM ¼ La C b Sg ð27Þ
4.1. Experimental results where L, C and S denote the comparison function of local
luminance, local contrast and structure, respectively. a 40, b 40
We compare our approach with standard RL, Wiener filtering, and g 40.
TV regularization and Levin’s method [22]. The parameters of each Since SNR and SSIM are used to measure the pixel value difference
method are tuned to ensure the best results. and structural difference, respectively; the data in table 1 show that

Fig. 4. Zoom of the images restored with different methods in the red rectangle, (a) blurred image, (b) Wiener filtering, (c) Levin’s method, (d) TV regularization, (e) standard RL
and (f) our method.
 W. Dong et al. / Optics & Laser Technology 43 (2011) 926–933 931

Fig. 5. Images restored with different methods: (a) blurred image and the PSF (23  23), the PSF is obtained with the method in [26], (b)–(f) zoom of the restored images in the
red square with Wiener filtering, Levin’ method, TV regularization, standard RL and our method. (For interpretation of the references to color in this figure legend, the reader is
referred to the web version of this article.)

Table 1 the RE curve of PLRL changes much smoother than that of RL, so the
Results of image assessment. convergence property of PLRL is better.
Fig. 6(e) shows the SNR curves of RL and PLRL within 200
Algorithms Wiener Levin’s TV RL Our
filtering method regularization method iterations. Clearly, the SNR for PLRL is higher than that of RL. The
highest SNR of RL is obtained after 29 iterations (SNR¼ 25.2398 dB)
SNR 16.4671 16.9954 16.9293 18.3503 19.1881 while for PLRL it is obtained after 48 iterations (SNR ¼25.6590 dB).
SSIM 0.5795 0.6469 0.6653 0.7149 0.8080 Fig. 6(b) and (c) shows the restored images by RL and PLRL after 200
iterations, we can see that Fig. 6(c) contains less noise.
From Eq. (23) and Eq. (25), we see that PLRL does not preserve
the global photometry as the method proposed in [34]. However,
our method acquires better results not only in pixel value, but also in experiments show that the loss is small. This is due to two factors.
structure than the other methods. One is that l is very small (which is tuned between 0.001 and 0.01)
and the values of the matrix f(Mon) are also not large. For example,
in Fig. 6, after 200 iterations, the maximum of f(Mon) is 1.9072, so
the values of 1 + lf(Mon) are no more than 1.0038 (In the experi-
5. Discussions ment, the parameters are l ¼ 0.002, t1 ¼0.2, scale ¼0.3). The other
factor is that f is piecewise, which contains a threshold t1, if an
The main problem of RL is that it does not converge, which element of Mon/255 is smaller than t1, the corresponding pixels in
makes it hard to decide when to stop the iteration. However, the on will be restored just the same as by RL. Fig. 6(f) shows the global
advantage of RL is the preservation of global photometry at each photometry ratio between the restored image and the blurred
iteration [34]. PLRL is derived from RL algorithm, in order to make a image for RL and PLRL, respectively. We can see that after 200
comparison, we design the following experiment. iterations, the ratio of PLRL is above 0.999. For the examples in
In Fig. 6(a), the clear image is blurred by a 7  7 Gaussian PSF and Figs. 3 and 5, the ratios of PLRL are 0.9785 and 0.9803, respectively.
the additive noise is Gaussian with u¼ 0, s ¼0.01. Both RL and PLRL So in most cases, the photometry loss of PLRL is acceptable.
are used to restore the blurred image. We use relative error (RE), What we must point out is that Eq. (22) is an essential boundary
which is defined in Eq. (28) [34] to compare the convergence process, whose role is just like the weighting function in [35]:
property of RL and PLRL. marking the discontinuities and giving them different smoothing
vffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi strength. Fig. 7(a) is the result of Mon (where on is show in Fig. 6(c)).
u
u:o o:2 We can see that the large discontinuities especially the edges are
RE ¼ t 2
ð28Þ
:o: strongly marked while the small discontinuities are not obviously
marked. This means if Mon is directly used without modification,
where o* is the restored image, o is the reference image. the large discontinuities such as the edges will be given large
Fig. 6(d) shows the RE curves of RL and PLRL within 200 iterations. smoothing strength and thus over smoothed, but the noise corre-
We can see that although both of them are not convergent, the RE of sponding to small discontinuities are given small smoothing and
PLRL is smaller than that of RL during the whole iterations. Moreover, preserved. Fig. 7(b) is the result of f(Mon), we can see that f nonlinearly
932 W. Dong et al. / Optics & Laser Technology 43 (2011) 926–933

Fig. 6. (a) Blurred image and the Gaussian PSF (7  7), the additive noise is Gaussian with u ¼0, s ¼ 0.01, (b) the restored image of RL after 200 iterations, (c) the restored image
of PLRL after 200 iterations, (d) curves of relative error (RE), (e) curves of SNR and (f) curves of global photometry ratio between the restored image and blurred image.

method with residual deconvolution to strengthen the image edges.

Visual experience and image assessment methods (SNR and SSIM)
show that the proposed method can work well when the obtained PSF
is not accurate and thus it is effective in restoring remote sensing
images.

Acknowledgements

We thank the reviewers for helping us to improve this paper.

This work is supported by State Key Development Program of Basic
Research of China (Grant no. 2009CB724006) and National Natural
Science Foundation of China (Grant no. 60977010).

Fig.7. (a) Result of Mon and (b) result of f(Mon).

References
adjusts the smoothing strength of Mon, the smoothing strength for the
noisy pixels on the background is enhanced while for edges it is [1] Tychonoff AN. On the stability of inverse problems. Doklady Akademii Nauk
SSSR 1943;39(5):195–8.
weakened. [2] Engl HW, Hanke M, Neubauer A. Regularization of inverse problems. Kluwer
The main difference between the weighting function in [35] and Academic Publishers; 2000.
f is: the weighting function is continuous and strictly deceasing on [3] Richardson WH. Bayesian-based iterative method of image restoration. Journal
of the Optical Society of America 1972;62(1):55–9.
[0,+ N) while f is increasing on [t1,+ N) (in most cases, scale Z0). So [4] Lucy LB. An iterative technique for the rectification of observed distribution.
the limitation of the weighting function is that the smallest discon- The Astronomical Journal 1974;79(6):745–54.
tinuities are given the largest smoothing strength, which may destroy [5] Hunt BR. Digital image processing. Proceeding of the IEEE 1975;63(4):
693–708.
the tiny textures of the restored image. For f(Mon), since disconti-
[6] Gonzalez RC, Woods RE. Digital image processing. 2nd ed.. Beijing: Publishing
nuities under the threshold t1 are not smoothed and small disconti- House of Electronics Industry; 2002.
nuities are given small smoothing strength, some tiny textures will be [7] Frieden BR. Restoring with maximum likelihood and maximum entropy.
Journal of the Optical Society of America 1972;62(4):511–8.
preserved. However, the main limitation of our algorithm is that the
[8] Dempster AP, Laird NM, Rubin DB. Maximum likelihood from incomplete data
penalty on large discontinuities is heavy, even though it is beneficial via the EM algorithm. Journal of the Royal Statistical Society, Series B
to suppress severe noise and ringing artifacts, it may also weaken the 1977;39(1):1–38.
image edges. [9] Zou M. Deconvolution and signal recovery. Beijing: National Defense Industry
Press; 2001.
[10] Zhang J, Zhang Q, He G. Blind image deconvolution by means of asymmetric
multiplicative iterative algorithm. Journal of the Optical Society of America
2008;25(3):710–7.
6. Conclusions
[11] Zhang J, Zhang Q, He G. Blind deconvolution: multiplicative iterative algorithm.
Optics Letters 2008;33(1):25–7.
In this paper, we propose a new regularized RL algorithm for [12] Zhu H, Lu Y, Wu Q. Blind image deconvolution subject to bandwidth and total
remote sensing image deconvolution. A piecewise local regulariza- variation constraints. Optics Letters 2007;32(17):2550–2.
[13] Fergus R, Singh B, Hertzmann A, Roweis ST, Freeman WT. Removing camera
tion term is introduced to make the standard RL algorithm capable shake from a single photograph. ACM Transactions on Graphics 2006;25(3):
of suppressing noise and ringing artifacts. We also combine our 787–794.
 W. Dong et al. / Optics & Laser Technology 43 (2011) 926–933 933

[14] Shan Q, Jia J, Agarwala A. High-quality motion deblurring from a single image. [25] Black MJ, Sapiro G, Marimont DH, Heeger D. Robust anisotropic diffusion. IEEE
ACM SIGGRAPH 2008. Article no: 73. Transactions on Image Processing 1998;7(3):421–32.
[15] Jia J. Single image motion deblurring using transparency. In: Proceedings of the [26] Yuan L, Sun J, Quan L, Shum H-Y. Image deblurring with blurred/noisy image
IEEE conference on computer vision and pattern recognition; 2007. p. 1–8. pairs. ACM SIGGRAPH 2007. Article no. 1.
[16] Dey N, Blanc-Feraud L, Zimmer C, Roux P, Kam Z, Olivo-Marin JC, Zerubia J. [27] Green PJ. On use of the EM algorithm for penalized likelihood estimation.
Richardson–Lucy algorithm with total variation regularization for 3D confo- Journal of Royal Statistical Society, Series B 1990;52(3):443–52.
cal microscope deconvolution. Microscopy Research Technique 2006;69(4): [28] Chellappa R, Chatterjee S. Classification of textures using Gaussian Markov
260–6. random fields. IEEE Transactions on Acoustics, Speech and Signal Processing
[17] Bioucas-Dias JM, Figueiredo MAT, Oliveira JP. Total variation-based image 1985;33(4):959–63.
deconvolution: a majorization-minimization approach. In: Proceedings of the [29] Besag J. Spatial interaction and the statistical analysis of lattice systems.
IEEE international conference on acoustics, speech and signal processing, vol. 2; Journal of the Royal Statistical Society. Series B (Methodological) 1974;36(2):
2006. p. 861–4. 192–236.
[18] Rodriguez P, Wohlberg B. Efficient minimization method for a generalized [30] Hammersley JM, Clifford P. Markov fields on finite graphs and lattices, 1971,
total variation functional. IEEE Transactions on Image Processing 2009;18(2):
/http://www.statslab.cam.ac.uk/ grg/books/hammfest/hamm-cliff.pdfS.
322–32.
[31] Patra SK, Mishra N, Chandrakanth R, Ramachandran R. Image quality improve-
[19] Zhang W, Zhao M, Wang Z. Adaptive wavelet-based deconvolution method for
ment through MTF compensation: a treatment to high resolution data. Indian
remote sensing imaging. Applied Optics 2009;48(24):4785–93.
Cartographer 2002;22:86–93.
[20] Neelamani R, Choi H, Baraniuk R. ForWaRD: fourier-wavelet regularized
[32] Ben-Ezra M, Nayar SK. Motion deblurring using hybrid imaging. In: Proceed-
deconvolution for ill-conditioned systems. IEEE Transactions on Signal Proces-
sing 2004;52(2):418–33. ings of the IEEE computer society conference on computer vision and pattern
[21] Neelamani R, Choi H, Baraniuk R. Wavelet-domain regularized deconvolution recognition, vol. 1; 2003. p. 657–64.
for ill-conditioned systems. In: Proceedings of the international conference on [33] Wang Z, Bovik AC, Sheikh HR, Simoncelli EP. Image quality assessment: from
image processing, vol. 1; 1999. p. 204–8. error visibility to structural similarity. IEEE Transactions on Image Processing
[22] Levin A, Fergus R, Durand F, Freeman WT. Image and depth from a conventional 2004;13(4):600–12.
camera with a coded aperture. ACM SIGGRAPH 2007. Article no: 70. [34] Bratsolis E, Sigelle M. A spatial regularization method preserving local
[23] Geman D, Yang C. Nonlinear image recovery with half-quadratic regulariza- photometry for Richardson–Lucy restoration. Astronomy and Astrophysics
tion. IEEE Transactions on Image Processing 1995;4(7):932–46. 2001;375(3):1120–8.
[24] Terzopoulos D. Regularization of inverse visual problems involving disconti- [35] Charbonnier P, Blanc-Feraud L, Aubert G, Barlaud M. Deterministic edge-
nuities. IEEE Transactions on Pattern Analysis and Machine Intelligence preserving regularization in computed imaging. IEEE Transactions on Image
1986;8(4):413–24. Processing 1997;6(2):298–311.