Markovian Random Fields and Comparison Between Different Convex Criterion

Optimization in Image Restoration

Jose Ismael de la Rosa , Member, IEEE, Jesus Villa, and Maria A. Araiza,
Signal Processing Laboratory - Engineering Faculty,
Universidad Autonoma de Zacatecas, Av. Lopez Velarde 801, Col. Centro,
98068 Zacatecas, Zac., MEXICO
Abstract
The present work illustrates some recent alternative methods to deal with digital image reconstruction. This collection of methods are inspired on the use of a class of Markov
chains best known as Markov Random Fields (MRF). All
of these new methodologies are also based on the prior
knowledge of some information which will permit more efficiently modeling the image acquisition process. The methods based on the MRFs are proposed and analyzed in a
Bayesian framework and their principal objective is to eliminate those effects caused by the excessive smoothness on
the reconstruction process of images which are rich in contours or edges. In order to respond to the edge preservation,
the use of certain convexity criteria are proposed which will
lead to obtain adequate weighting of cost functions (halfquadratic) in cases where discontinuities are remarked and,
even better, for cases where such discontinuities are very
smooth. The final aim is to apply these methods to problems in optical instrumentation.

1. Introduction
The use of powerful methods proposed in the seventies (iterated conditional modes) [2, 3, 9], are nowadays essential at least in the cases of image segmentation and image
restoration [1]. The basic idea of these methods is to construct a Maximum a posteriori (MAP) of the modes or so
called estimator of true images by using Markov Random
Fields (MRF) in a Bayesian framework. The evolution of
the basic idea has caused the development of new algorithms which consider new models of contextual information which is lead by the MRFs and the final aim is the
restoration of real images (practical data). The idea is based
in a robust scheme which could be adapted to reject out

Corresponding autor,e-mail:ismaelrv@ieee.org

liers, tackling situations where noise is present in different

forms during the acquisition process.
The image restoration approaches or recuperation of an
image to its original condition given a degraded image,
passes by reverting the effects caused by a distortion functional which must be estimated. In fact, the degradation
characteristics is a crucial information and it must be supposed known or estimated during the inversion procedure.
Typically this is a point spread function (PSF) from the distortion which can be linked with the probability distribution
of the noise contamination, in the case of MAP filters, usually the additive Gaussian noise is considered. There is another source of information which imposes a key rule in the
image processing context, this is the contextual or spatial information, that represents the likelihood or correlation between the intensity values of a neighborhood of pixels well
specified. The modelling when using MRF take into account
such spatial interaction and it was introduced and formalized in [2] where it is shown the powerfulness of these statistical tools [3, 4, 5, 9, 20]. Combining both kinds of information in an statistical framework, the restoration is lead by
an estimation procedure given the maximum a posteriori of
the true images when the distortion functionals are known.
The implemented algorithms were developed considering a
slightly degraded signal, where the resulting non-linear recursive filters show excellent characteristics to preserve all
the details contained in the image, and on the other hand,
they smooth the noise components.
The section 2 describes the general definition of an MRF
and the proposal of the MAP estimator. The potential functions must be obtained or proposed to conduct adequately
the inversion process, such functions are described in section 3 where the convexity is the key to formulate an adequate criterion to be minimized. In sections 4 and 5 are discussed briefly the MAP estimators resulting from different
MRF structures and some illustrative results. Finally in section 6 are given some partial conclusions and comments.

17th International Conference on Electronics, Communications and Computers (CONIELECOMP'07)

0-7695-2799-X/07 $20.00 2007

2. Markov random fields and MAP estimation

The problem of image estimation (e.g. restoration) into a
Bayesian framework deals with the solution of an inverse
problem, where the estimation process is carried out in a
whole stochastic environment. All variables presented along
the text are, x: which represent a Markov random field (or
image to be estimated), y: represents the observed image
with noise and distorted, and x
b: is the estimator of x with
respect to data y, and p() is a probability density function.
The most popular estimators used nowadays are:
Maximum Likelihood estimator (ML): this
estimator produces excessive noise. An ill posed problem is presented when the quantity of data is poor.
Of course, it is important to exploit all known information or so called prior information about any
process under study which gives a better estimator called
Maximum A Posteriori (MAP) estimator:
x
bMAP

= arg min p(x|y)

xX

= arg max (log p(y|x) + log g(x)) ,

(1)

xX

in this case, the estimator is regularized by using a

Markov random field function g(x) which model all
prior information as a whole probability distribution,
where X is the set of pixels x capable to minimize
p(x|y), and p(y|x) is the likelihood function from y
given x.
The Markov random fields (MRF) can be represented in a
generic way by using the following equation:

!
X
1
Vc (x) ,
(2)
g(x) = exp
Z
cC

where Z is a normalization constant, C is a set of cliques

c or local neighborhood of pixels, and Vc (x) is a weighting function given over the local neighborhood. Generally,
the cliques correspond to the set of neighborhoods of pixels. An introduced theorem by Hammersley-Clifford [2, 9]
probes the equivalence between the Gibbs distribution and
the MRFs. The Markov random fields have the capability to
represent various images sources. The main disadvantage of
the use of MRFs is that the estimation procedure is lead under the local minimization schemes (overall grows the time
of computation), and the proposal of global minimization
schemes has been recently introduced by modifying the local structure of the MRFs. There is a variety of MRF models which depend on the cost function also known potential
functions that can be used. Each potential function characterizes the interactions between pixels in the same local

group. As an example, the following family represent convex functions:

X
([xi xj ]),
(3)
{i,j}C

where is a constat parameter to be selected, and x belong

to the local group of pixels.

3. Generalizaed GMRF and half-quadratic

functions
In recent woks [5][14] some new potential functions were
introduced, such proposed functions are semi-quadratic
functionals or half-quadratic and them characterize certain convexity into the penalization term [10, 11](eg.
extension to penalization) which permit to build efficient and robust estimators in the sense of data preservation which is linked to the original or source image, also
the the time processing decreases with respect to other proposed schemes. On the other hand, in previous works
from Alison Gibbs [13] it has been proposed to obtain the posterior distribution of the images, in this case
it is necessary to use of sophisticated stochastic simulation techniques based on the Markov Chain Monte Carlo
(MCMC) [15, 19]. If is possible to obtain the posterior distribution of any image, thus, it is possible to sample
from such posterior distribution and obtain the MAP estimator, or other estimators such as the median estimator
which is sensible to be equal to the MAP estimator under certain circumstances imposed by the noise structure,
the MAP and the median estimators searches the principal mode of the posterior distribution.
In the present paper some potential functions are compared, the proposed generalized Guassian MRF introduced
by Bouman in [4] is compared with respect to semi-Huber
[8], Welch, and Tukey potential functions, these two last
functions were used in recent works of Rivera [16][18]
proving excellent performance.

3.1. Generalizaed GMRF

If one consider to generalize the Gaussian MRF (when p =
q = 2 in equation (16), one has a GMRF) as proposed by
Bouman, one can redefine
= [xi xj ].

(4)

where the generalized potential functions can be changed as

17th International Conference on Electronics, Communications and Computers (CONIELECOMP'07)

0-7695-2799-X/07 $20.00 2007

() = ||p ,

1<p<2

(5)

obtaining the GGMRF

X
log g(x) = p
as xps +
sS

This function is also half-quadratic such as the Tukey function.

bsr |xs xr |p +cte,

{s,r}C

(6)
where cte is a constant, and the weighting parameters as >
0 and bsr > 0. In practice it is recommended to take as = 0
thus, the unicity of x
bMAP , can be assured by

X
log g(x) = p
bsr |xs xr |p + cte, (7)

3.4. Tukey potential function

This is another hard redescender potential function, also
used by Rivera in [16]

X
bsr 1 (x)
log g(x) =
{s,r}C

{s,r}C

+(1 )

bsr 3 (x) + cte,

log p(y|x) is strictly convex and so x

bMAP is continuous in
y, and with respect to the power p. The choice of p is capital, since the selection of p constrains the convergence speed
of the global estimator, and the quality of the restored image.

where

3.2. Semi-Huber potential function

and where k is also a scale parameter.

{s,r}C

(8)
where
2
1 (x) = 0
2

!
41 (x)
1+
1 ,
20

(9)

where 0 > 0 is a constant value. On the other hand, 1 (x)

can be the quadratic function described by 1 (x) = e2 ,
where e = (xs xr ). The potential function can also be
considered as half-quadratic (HQ) functional.

3.3. Welch potential function

Known as a hard redescender potential function, and used
by Rivera in [16]

X
bsr 1 (x)
log g(x) =
{s,r}C

+(1 )

(10)

bsr 2 (x) + cte,

{s,r}C

where k is a scale parameter and

2 (x) = 1

{s,r}C

3 (x) =

1 (1 (2e/k)2 )3 ,
1,

for |e/k| < 1/2,

otherwise.
(13)

To assure completely the robustness diminishing at the same

time the convergence speed, the Huberlike norm or semi
Huber potential function can be used as described in [8],
this proposal is adjusted in this work in two dimensions by
the following equation (SH):

X
X
bsr 1 (x) +
bsr 1 (x)+cte,
log g(x) =
{s,r}C

(12)

1
exp(k1 (x))
2k

(11)

4. MAP estimators and practical convergence

In this section some estimators are deduced, the single problem of filtering noise to restore an observed signal Y lead to
establish the estimators, the observation equation could be
Y = X + n Z, where Z is Gaussian noise.

(14)

The MAP estimator for this particular problem when using

the GGMRF is given by,
(
X
x
bMAP1 = arg min
|ys xs |q
xX
sS

(15)

X
q p
p
,
+
bsr |xs xr |

{s,r}C

where the power q is given according with the hypothesis

made about the noise contamination (q = 2 for Gaussian
noise), and bsr is chosen according to the number of
neighbors. Thus, the minimization problem leads to consider various methods:
global iterative techniques such as: the descendent gradient, conjugate gradient, Gauss-Seidel, Over-relaxed
methods, etc.
or local minimization techniques: minimization at each
pixel xs (which generally needs more time, but are
more precise).
Local techniques were used in this work, where the expectation maximization (EM) was not implemented, since all
parameters included into the potential functions were chosen heuristically or according to values proposed in some

17th International Conference on Electronics, Communications and Computers (CONIELECOMP'07)

0-7695-2799-X/07 $20.00 2007

references. For example, the local MAP estimator for the

GGMRF is given by
(
q

q p

x
bs = arg min |ys xs | +
xX

)
brs |xs xr |

rs

(16)
where according to the value of parameters p and q the performance of such estimator varies. For example if p = q =
2, the obtained estimator is similar to the least squares one
since the likelihood function is quadratic, with an additional
term of penalization which degrades the estimated image
P
ys + ()2 rs brs xr
P
.
x
bs =
1 + ()2 rs brs

(17)

On the other hand, in the case of p = q = 1, the criterion is

absolute, and the estimator converges to the median estimator which in practice is difficult to implement
x
bs = median (ys , xr1 , . . . , xrI ) ,

(18)

this criterion is not differentiable and this fact causes instability in the minimization procedure. For intermediate
values of p and q the estimators become sub-optimal, and
the use of iterated methods can be used to minimize the
obtained criterions, such iterative methods are the sub
gradient, or the LevenbegMarquardt method of MATLAB
7 which was used in this work. The local or global condition of the estimator depends on
1) if one has values of 1 < p < 2: the estimator x
bmin.loc
x
bmin.glob , which means that a local minimum would
coincide with a global minimum,
2) moreover, if p 6= q, the criterion is not homogeneous,
but: x
b(y, ) = b
x(y, 1q/p ), assuring the convergence and existence of the estimator which is continuous with respect to p.
A second MAP estimator can be obtained when using the
semiHuber potential, the global estimator can be described
by the equation

x
bMAP2 = arg min

xX

1 (x) +

|ys xs |2 +

sS

X
{s,r}C

bsr

{s,r}C

bsr 1 (x)

(19)
as in the previous estimator, it has been proposed to implement the local estimator which leads to a similar expression
as the equation (16) for the first local MAP estimator.

The third MAP estimator is obtained when using the

Welch potential function, that is,

X
X
|ys xs |2 +
bsr
x
bMAP3 = arg min
xX
sS
{s,r}C

X
1 (x) + (1 )
bsr 2 (x) .

{s,r}C

(20)
And finally, the four MAP estimator is deduced from the
Tukey potential function, giving the following global estimator

X
X
|ys xs |2 +
bsr
x
bMAP4 = arg min
xX
sS
{s,r}C

X
1 (x) + (1 )
bsr 3 (x) .

{s,r}C

(21)
The use of a prior distribution function based on the
logarithm, with any degree of convexity and quasihomogeneous permits to consider a variety of possible
choices of potential functions. May be, the most important challenges that must be well resolved are: the adequate selection of hyper-parameters from potential functions, where different versions of the EM algorithms try to
tackle this problem [5, 7], another is the minimization procedure which in any sense will regulate the convergence
speed as proposed in [11, 16].

5. Denoising experiments
Continuing with the problem of filtering noise some estimation results are presented when images are only contaminated by Gaussian noise, and there are no other type of distortions, the first experiment was made considering the next
model
Y = X + n Z,

Z N (0, In2 ),

n = 2,

where I is the identity matrix. The results are compared using different values for p and preserving q = 2, some images where used to probe the performance of the presented
estimators. Here are presented some results based on the
analysis of the standard image of Lena, different levels
of noise were added to the image: Z N (0, In2 ), with
n = 4 this level of noise does not degrade visually the image (see Figure 1), but increasing the value of n = 8 the
obtained degradation is perceptible and difficult to eliminate (see Figure 2) where the performance of the MAP1 estimator depends on the choice of p and also the level of
noise. In the case of Figure 3 some visual results when using the other three MAP estimators are given, the performance of the MAP2 is equal or better than the MAP1, but

17th International Conference on Electronics, Communications and Computers (CONIELECOMP'07)

0-7695-2799-X/07 $20.00 2007

the time processing is far better, since for best results obtained for MAP1 were when p = 1.1 and the time was approximately 1200 seconds ,while for the MAP2, when using a value for 0 = 15, the computation time was 172
seconds (the image dimensions are 128 128 pixels). The
other two estimators have a good performance in the sense
of the computation time, but the tuning of hyper-parameters
is a drawback which may be attended to obtain better performance in the sense of restoration quality. For example, in
Figure 3 the MAP3 estimation results seems worst than the
MAP2, for this case the next values = 15, k = 1000 and
= 0.045 were used into the potential function and the obtained time of computation was 155 seconds. For the case
of the Tukey potential function, the same problem as the
tuning problem presented for the Welsh potential must be
solved to obtain better performance as depicted by Rivera
in [16], the obtained results presented here consider for the
MAP4 estimator the next values for = 35, k = 1000 and
= 0.05. This last time of computation was of 198 seconds.
In the case of the first estimator (MAP1) obviously the
performance depends on the choice of p and q, the best performance in the sense of the restoration quality is presented
when p 1, but unfortunately, the time of computation
grows exponentially. On the other hand, the use of halfquadratic potential functions permits more flexibility for the
time of computation, but still is a challenge to tune correctly
the hyper-parameters in order to obtain better performance
in the sense of restoration quality, may be the most simple
potential function to tune is the semi-Huber, but our experience over this function is reflected in a large use of it for
robust estimation. In the case of the Welsh and Tukey functionals the tuning problems must be resolved implementing
in correct ways more sophisticated algorithms based on the
expectation maximization method, this last task remains as
future implementation, that means that this work is placed
as one of our objectives. Some interesting applications of
robust estimation are particulary focused in phase recovery
from fringe patterns as presented in recent work [21], and
phase unwrapping, in this sense some filtering results were
also obtained using the MAP presented estimators.

6. Conclusions and comments

Some advantages on the use of GGMRF are: the continuity of the estimator is assured in function of the data values
when 1 < p 2. The edge preserving is also assured, over
all when p 1 and obviously depends on the choice between the interval 1 < p < 2. The robustness is also assured in the same sense. The use of semi-quadratic or halfquadratic potential functions also present some advantages:
the convexity of this functions is relaxed with respect to
the GGMRF, this fact gives as result the decrement of the

a)

b)

20

20

40

40

60

60

80

80

100

100

120

120
20

40

60

80

100

120

20

40

c)

60

80

100

120

80

100

120

d)

20

20

40

40

60

60

80

80

100

100

120

120
20

40

60

80

100

120

20

40

60

Figure 1. a) Lena original image (128 128), b)

Image with low level noise, c) MAP1 estimation with p = 1.5 and q = 2, d) MAP1 estimation with p = 1.2 and q = 2, = 10.
time of computation, which is a good advantage over the
GGMRF, but the tuning of hyper-parameters is most complicated, since one has more degrees of freedom. In the case
of the semi-Huber potential function the tuning is less complicated and of course, the estimator manipulation is far
simplest than the Welsh and Tukey, however this problem
can be solved as argued by Idier [5] and Rivera [16] implementing more sophisticated algorithms. Moreover, the
advantages presented by those estimators could overcome
the disadvantages experimenting largely with them, and establishing the aim to implement more sophisticated algorithms. The finality is to build a serial of software tools for
image analysis focused for instance to optical instrumentation tasks such as those treated by the work of Villa in [21].

Acknowledgements
Many thanks to PROMEP of Mexico, this work was partially supported by the Mexican Program for Professors
Technical Improvement (PROMEP) under Grant UAZAC
PTC 24-103.5/03/1127.

References
[1] H. C. Andrews, and B. R. Hunt, Digital image restoration,
New Jersey, Prentice-Hall, Inc., 1977.
[2] J. E. Besag, Spatial interaction and the statistical analysis
of lattice systems, J. Royal Stat. Soc. Ser. B, Vol. B-36, pp.
192236, 1974.
[3] J. E. Besag, On the statistical analysis of dirty pictures, J.
Royal Stat. Soc. Ser. B, Vol. B-48, pp. 259302, 1986.

17th International Conference on Electronics, Communications and Computers (CONIELECOMP'07)

0-7695-2799-X/07 $20.00 2007

[4] C. Bouman, and K. Sauer, A Generalizaed Gaussian Image

Model for Edge-Preserving MAP Estimation, IEEE Trans.
on Nuclear Science, Vol. 99, No. 4, pp. 11441152, 1992.
[5] K. Champagnat, and J. Idier, A Conection Between HalfQuadratic Criteria and EM Algorithms, IEEE Signal
Processing Letters, Vol. 11, No. 9, pp. 709712, Sept. 2004.
[6] P. Ciuciu, J. Idier, and J.-F. Giovannelli, Regularized Estimation of Mixed Spectra Using Circular Gibbs-Markov

[7]

[8]

[9]
a)

b)

20

20

40

40

60

60

80

80

100

100

120

[10]

[11]

120
20

40

60

80

100

120

20

40

c)

60

80

100

120

d)

20

20

40

40

60

60

80

80

100

100

120

[12]

[13]

120
20

40

60

80

100

120

20

40

60

80

100

120

Figure 2. a) Lena original image, b) MAP1 estimation with p = 1.1 and q = 2, c) MAP1 estimation with p = 1.2 and q = 2, d) MAP1 estimation with p = 1.5 and q = 2, = 5, and
n = 8 .

[14]

[15]

[16]
a)

b)

20

20

40

40

60

60

80

80

100

100

120

[17]

[18]

120
20

40

60

80

100

120

20

40

c)

60

80

100

120

[19]

d)

20

20

40

40

60

60

80

80

100

100

120

[20]

[21]

120
20

40

60

80

100

120

20

40

60

80

100

120

Model, IEEE Trans. on Signal Processing, Vol. 49, No. 10,

pp. 22022213, Oct. 2001.
P. Ciuciu, and J. Idier, A half-quadratic block-coordinate
descent method for spectral estimation, Journal of Signal
Processing, Vol. 82, pp. 941959, 2002.
J. I. De la Rosa, and G. Fleury, Bootstrap methods for
a measurement estimation problem, IEEE Trans. Instrum.
Meas., Vol. 55, No. 3, pp. 820827, June 2006.
S. Geman, and C. Geman, Stochastic relaxation, Gibbs distribution, and the Bayesian restoration of images, IEEE
Trans. Pattern Anal. Machine Intell., Vol. PAMI-6, pp. 721
741, Nov. 1984.
S. Geman, and G. Reinolds, Constrained restoration and the
recovery of discontinuities, IEEE Trans. Pattern Anal. Machine Intell., Vol. 14, pp. 367383, Mar. 1992.
S. Geman, and C. Yang, Nonlinear image recovery with
half-quadratic regularization, IEEE Trans. Image Processing, Vol. 4, pp. 932946, July 1995.
J.-F. Giovannelli, J. Idier, R. Boubertakh, and A. Herment,
Unsupervised Frequency Tracking Beyond the Nyquist Frequency Using Markov Chains, IEEE Trans. on Signal
Processing, Vol. 50, No. 12, pp. 29052914, Dec. 2002.
A. L. Gibbs, Convergence of Markov Chain Monte Carlo algorithms with applications to image restoration, Ph.D. Thesis, Department of Statistics, University of Toronto, URL :
www.utstat.toronto.edu, 2000.
J. Idier, Convex Half-Quadratic Criteria and Interacting
Auxiliary Variables for Image Restoration, IEEE Trans. on
Image Processing, Vol. 10, No. 7, pp. 10011009, July 2001.
R. M. Neal, Probabilistic inference using Markov Chain
Monte Carlo methods, Tech. Rep., CRG-TR-93-1, Department of Computer Science, University of Toronto, URL :
www.cs.toronto.edu/radford, 1993.
M. Rivera, and J. L. Marroquin, Efficent half-quadratic regularization with granularity control, Image anv Vision Computing, Vol. 21, pp. 345357, 2003.
M. Rivera, and J. L. Marroquin, Half-quadratic cost functions for phase unwrapping, Optics Letters, Vol. 29, No. 5,
pp. 504506, 2004.
M. Rivera, Robust phase demodulation of interferograms
with open or closed fringes, J. Opt. Soc. Am. A, Vol. 22, No.
6, pp. 11701175, 2005.
C. P. Robert, and G. Casella, Monte Carlo Statistical Methods, Springer Verlag, 2nd Edition 2004.
K. Sauer, and C. Bouman, Bayesian Estimation of Transmission Tomograms Using Segmentation Based Optimization, IEEE Trans. on Image Processing, Vol. 2, No. 3, pp.
296310, July 1993.
J. J. Villa, J. I. De la Rosa, G. Miramontes, and J. A. Quiroga,
Phase recovery from a single fringe pattern using an orientational vector field regularized estimator, J. Opt. Soc. Am.
A, Vol. 22, No. 12, pp. 27662773, 2005.

Figure 3. a) Image with noise, n = 8, b) MAP2

Estimation, c) MAP3 estimation, and d) MAP4
estimation.

17th International Conference on Electronics, Communications and Computers (CONIELECOMP'07)

0-7695-2799-X/07 $20.00 2007