856
6, JUNE 1995
[5] C. Bouman and B.Liu, “Multiple resolution segmentation of textured
images,” IEEE Trans. Pattern Anal. Machine Intell., vol. 13, no. 2, pp.
99-113, Feb. 1991.
[6] B. S. Manjunath and R. Chellapa, “A unified approach to boundary
perception: Edges, textures, and illusory contours,” IEEE Trans. Neural
Networks, vol. 4 no. 1, pp. 96-108, Jan. 1993.
[7] T. Pavlidis and Y.-T. Liow, “Integrated region growing and Edge
detection,” IEEE Trans. Pattern Anal. Machine Intell., vol. 12, no. 3,
pp. 225-233, Mar. 1990.
[8] K.-B. Eom and R. L. Kashyap, ‘Texture and intensity edge detection
with random field models,” in Proc. IEEE Comp. Soc. Workshop Comp.
vision, Miami Beach, FL, Nov. 30-Dec. 2, 1987, pp. 29-34.
[9] Y. Xiaohan et al., “A new algorithm for texture segmentation based on
edge detection,” Pattern Recognition, vol. 24, no. 11, pp. 1105-1112,
Nov. 1991.
[lo] D. He and L. Wang, “Detecting texture edges from images,’’ Pattern
Recognition, vol. 24, no. 6, pp. 595-600, June 1992.
[ 111 T. W. Ryan, P. J. Yuen, and B. R.Hunt, “Extraction of shoreline features
by neural nets and image processing,” Photogram. Eng. & Remote
Sensing, vol. 57, no. 7, pp. 947-955, July 1991.
[12] S. R. Yhann, “A multi-resolution approach to texture analysis and seg-
mentation using neural network,” Ph.D. dissertation, Dept. of Electrical
and Computer Engineering, University of Miami, E, 1992.
[13] S. R. Yhann and T. Y. Young, “A multi-resolution approach to texture
segmentation using neural networks,” in Proc. 10th Int. Con$ Pattern
Recog., Atlantic City, NJ, June 1990, pp. 513-517.
[14] G. S. Robinson, “Detection and coding of edges using directional
masks,” in Proc. SPIE Con$ Advances Image Transmission Techniques,
San Diego, CA, Aug. 1976.
[15] R. Nevatia, “Locating object boundaries in textured environments,”
IEEE Trans. Comput., vol. C-25, no. 11, pp. 1170-1175, Nov. 1976.
[16] C. R. Giardina and E. R. Dougherty, Morphological Methods in Image
and Signal Processing. Englewoods Cliffs, NJ: Prentice Hall, 1988.
[17] P. Brodatz, Textures-A Photographic Album for Artists and Designers.
New York: Dover, 1966.
[I81 D. Hanvood, M. Subbarao, and L. S. Davis, “Texture classification by
local rank correlation,” Comp. vision, Graph., Image Processing, vol.
32, pp. 4-11, 1985.
[19] M. Unser, “Local linear transforms for texture measurements,” Signal
Processing, vol. 11, no. 1, pp. 61-74, 1986.
[20] C. C. Gotlieb and H. E. Kreyszig, ‘Texture descriptors based on co-
occurrence matrices,” Comp. vision, Graph., Image Proc., vol. 51, pp.
70-86, 1990.
[21] H. Greenspan, R. Goodman, and R. Chellappa, “Texture analysis via
unsupervised and supervised learning,” in Proc. Int. Joint Con$ Neural
Networks, Seattle, WA, July 1991, pp. 639-644.
[22] B. P. Kjell and C. R. Dyer, “Segmentation of textured images,” in Proc.
IEEE Con$ Comput. vision Pattern Recognition, Miami Beach, FL, June
1986, pp. 47W81.
[23] R. Wang et al., ‘Texture analysis based on local standard deviation
of intensity,” in Proc. IEEE Con$ Comp. vision Pattern Recognition,
Miami Beach, FL, June 1986, pp. 482-488.
1057-7149/95$04.00 0 1995 IEEE
IEEE TRANSACTIONS ON IMAGE PROCESSING, VOL. 4, NO.
A Markov Random Field Model-Based Approach
to Unsupervised Texture Segmentation
Using Local and Global Spatial Statistics
Charles Kervrann and Fabrice Heitz
Absfract-Many studies have proven that statistical model-based tex-
ture segmentation algorithms yield good results provided that the model
parameters and the number of regions be known a priori. In this
correspondence,we present an Unsupervised texture segmentationmethod
that does not require knowledge about the Merent texture regions, their
parameters, or the number of available texture classes. The proposed
algorithm relies on the analysis of local and global second and higher
order spatial statistics of the original images. The segmentation map is
modeled using an augmented-state Markov random field, including an
outlier class that enables dynamic creation of new regions during the
optimization process. A Bayesian estimate of this map is computed using a
deterministic relaxation algorithm. Results on real-world textured images
are presented.
I. INTRODUCTION
unsupervised image segmentation remains a fundamental issue in
low-level image analysis and computer vision since most high-level
processing schemes must rely on reliable and robust segmentations. In
this correspondence,we address the problem of nonsupervised texture
segmentation. Although this problem has received considerable atten-
tion in the last decade 141,181, [lo], [ll], [13], completely data-driven
segmentation methods, able to handle real-world textured images,
remain beyond the present state-of-the-art. The algorithm presented
in this correspondence is unsupervised in the following sense:
it does not require a priori knowledge on the textures of the
image (in particular no learning step is necessary);
the number of regions or texture classes needs not to be known
either.
Early attempts at nonsupervised texture segmentation have made
use of texture measures associated with edge detection techniques
[8]. A more powerful approach consists in modeling the different
textures with stationary statistical models, for instance Gauss-Markov
random fields (GMRF) [lo]-[12]. The image is usually divided into
disjoint local windows [lo]-[12]. Each region is assumed to contain
mainly one texture. This assumption enables one to identify locally
the parameters of stationary texture models [ll], [12]. Windows
with nearby parameters are then merged using clustering techniques
[IO], [12]. One problem with these methods is the difficulty of
the computation, from local windows, of reliable parameters for
the underlying model. Several authors [4], [7] have $us proposed
describing texture using statistical features [6] that are computed in
local windows. The use of local texture features, instead of global
texture models, reduces significantly the number of parameters to be
determined. These local texture features are often associated with a
global representation of the image partition using nonlinear Markov
random field (MRF) models 131, [4], [Ill.
Manuscript received July 28, 1993; revised July 2, 1994. This work was
supported by the GDR TdSI 134 and by MRE and CNRS in the context of
the PRC Program “Man Machine Interface.” The associate editor coordinating
the review of this paper and approving it for publication was Prof. Roland T.
Chin.
The authors are with IRISA/INRIA, Campus Universitaire de Beaulieu,
35042 Rennes Cedex, France.
IEEE Log Number 9411143.
IEEE TRANSACTIONS ON IMAGE PROCESSING, VOL. 4, NO. 6, JUNE 1995
global candidate region R(e,)
Fig. 1. Sets of texture features supporting the local and global
distributions.
before relabeling : R = 4 after relabeling : R = 6
Fig. 2. Relabeling of the outlier class p. Label p is maintained on small
connected components (of size less than 1% of the image size). New regions
are created for all other connected components.
An intricate issue for these different approaches is the determina-
tion of the number of textures present in the image. Model fitting
criteria have been proposed recently to address this problem [13].
In this correspondence, we propose an alternate approach relying
on an augmented state-MRF model that enables one to handle
an a priori unknown number of regions'. Our model is inspired
by the model described by Geman et al., in [4], with several
significant differences. In [4], various texture features are used, along
with an MRF model of the segmentation map. The algorithm is
partially unsupervised but the number of regions is assumed to be
known approximately and thresholds for discriminating the different
textures have to be fine-tuned. In the approach proposed here, the
segmentation is obtained using a deterministic relaxation procedure
that simultaneously determines the current number of regions. This
is achieved by introducing an additional state (corresponding to an
outlier class) in the MRF model representing the image partition.
This additional label enables the dynamic creation of new regions
during the relaxation process (a similar idea has been suggested for
motion-based segmentation by Bouthemy and FranGois [2]). Mixed
windows may thus be split in several sub-regions, leading to reliable
and robust parameter estimation. Besides, the method only q u i r e s
the tuning of one parameter, which is a probability that determines
all the thresholds on the texture features.
The remainder of this correspondence is organized as follows. In
Section II, we briefly describe the general framework of the statistical
approach used here (global Bayesian estimation, MRF models, and
This correspondence is a short version of the technical report [7] in which
additional details about the method as well as other experimental results can
be found.
857
energy minimization techniques). The segmentation algorithm is
described in Section III. Results on mosaics of natural textures, as
well as on real-world textured images are discussed in Section N.
n. BAYESIAN
IMAGE ANALYSE AND MARKOV RANDOM FIELDS
The general mathematical framework under concern in this cor-
respondence is based on global Bayesian estimation along with
MRF models. The goal of this theory is to extract a 2-D field
of discrete labels describing the information of interest in a given
image processing task from a given set of observed variables (in
texture segmentation, for instance, the labels to be estimated are
region labels). Let 0 = (Os, s E S) designate an observation field
defined on a rectangular lattice S. k t E = (Es, s E S) denote
the unobserved (hidden) label field, defined on the same lattice S.
Realizations of fields 0 (resp. E) will be denoted o = {os, s E S }
(resp. e = {e,, s E S}).Let A be the (discrete) state space of random
variable e, and Cl the (finite) set of all possible label configurations
e = {e,, s E S}. If (E, 0) is assumed to be an MRF with respect
to a chosen neighborhood system D = {Gs,s E S}, it can be shown
[5] that the best estimate i of e given o (according to the maximum
a posteriori (MAP) criterion) is obtained by minimizing a global
energy function U ( o , e):
d = arg min U ( o , e). (1)
eER
Due to the Markovian property of the field, the energy decomposes
as follows [5]:
~ ( 0e ), = C v,(o, e) (2)
CEC
where C denotes the set of cliques associated to neighborhood system
0. Cliques c are subsets of sites that are mutual neighbors. The
potential function V, is locally defined on clique c and expresses
the local interactions between the different variables of the clique
(see [3] and [5] for an extensive presentation of MRF models).
Defining global energy functions is a powerful tool for specifying
nonlinear interactions between different image features (luminance,
edges, region labels, etc.). They help to combine and organize spatial
and temporal information by introducing strong generic knowledge
about the features to be estimated. Energy minimization techniques
have been used in a wide range of applications including image
restoration, edge detection, luminance and texture segmentation, sur-
face reconstruction, visual motion analysis, and scene interpretation
[3]. The choice of particular energy functions is either heuristic or
may be guided by a statistical modeling of the interaction between
the variables, as will be seen in the following.
m. UNSUPERVISED TEXTURE SEGMENTATION
A. Statistical Texture Features
The observations, on which the segmentation process is based,
correspond to vectors of texture features describing various spatial
statistics. These features are computed on local windows and are
assigned to the pixel located at the center of the window. To each
site is thus assigned a feature vector os = [ ~ ' ( s ) , . . . , om(s)] where
oa(s), i = 1, . . . , m designates the ith texture feature computed at
site s and m is the number of texture features. The features that have
been used here include seven features proposed by Haralick (mean,
variance, energy, correlation, entropy, contrast, and homogeneity),
derived from the joint occurrency matrices [6] and three other features
proposed by Geman et aL, [4] (the gray level, the local contrast, and
an "isotropic residual"). Several other parameters (all the parameters
defined in [6] and features based on the Fourier transform) have
858
Fig. 3. (a) Three-region textured image. (b) Ground truth (the boundaries of the different regions are superimposed on the original image (c) Result of the
unsupervised segmentation procedure ( p = 0.75). (d) Difference between the ground truth and the unsupervised segmentation.
been tested, but have not been included in the final version of the
segmentation algorithm. The set of texture measures retained here
has led to satisfactory results on all the processed images'.
B. The Global Energy Function
As in Section II, e,, s E S will designate the hidden variable
assigned to site s, which describes the label of the region to which
site s belongs. o = {os, s E S} represents the observation field
corresponding to vectors of statistical texture features. The issue is to
estimate field e = {es, s E S} and to determine R (the number of
regions in the image) from the observations o. A MAP estimator is
used to this end. The MAP estimate of e is obtained by minimizing
the following global energy function:
d = arg min o, R)+ Uz(e).
Ul(e, (3)
eEQ
This energy function is defined as the sum of two contributions. The
term U2 describes the prior on the region label map e. This tenn
is related to a Markov model associated to an eight-neighborhood
system. The model favors spatially homogeneous regions, by the
'Techniques for selecting "optimal" texture features for a particular class
of images, within a supervised framework, have been proposed recently in
[9]. This issue, however, has not been addressed here.
EEE TRANSACTIONS ON IMAGE PROCESSING, VOL. 4, NO. 6, JUNE 1995
choice of potential functions of the form
where p is a positive parameter and 6(z) = 1 for 2 = 0, 6(z) = 0
otherwise. In this energy function, only two-element cliques are
considered [7].
Term U1 describes the interaction between observations and labels.
It is inspired by the model described by Geman et al., [4], with several
significant differences. In [4], the model is based on a Kolmogorov-
Smirnov (KS) distance expressing the similarity of the distribution
of texture features in blocks centered at two neighboring sites s and
t. We introduce here a different potential function that measures the
distance, at a given site s, between the distribution of texture features
in a block B, centered at s (local distribution) and the distribution of
texture features in the whole candidate region to which we want to
assign site s (global distribution) (Fig. 1). This technique enables us,
in particular, to create new regions by detecting outliers during the
relaxation process, as will be explained in Section HI-C. Accordingly,
energy function U1 is decomposed as
e, R) =
u~(o, V ~ { O ( Bo[~(e~)I)
~), (5)
SES
IEEE TRANSACTIONS ON IMAGE PROCESSING, VOL. 4, NO. 6, JUNE 1995
Fig. 4. (a) Three-region textured image. (b) Ground truth (the boundaries of the different regions are superimposed on the original image (c) Result of the
unsupervised segmentation procedure ( p = 0.65). (d) Difference between the ground truth and the unsupervised segmentation.
where o(B,) = {or, r E B,} is the set of feature vectors on block
B, centered at s and o[R(e,)] = {or, r E R(e,)} is the set of
feature vectors belonging to the region labeled e, (Fig. 1).
The potential function VI is defined by
m statistics. The minimization of the global energy function (3), as well
) , ( B , ) )>c(")I-11
V ~ { O ( B ,o) [, ~ ( e , ) l ) ={~~ [ d ( o ( ' ) ( ~ ( e , )o(*),
2=1
(6)
where d{o("[R(e,)], O ( ~ ) ( Bstands~ ) } for the KS distance between
the distribution of feature o(') on the whole region labeled e,, and
the distribution of the same feature on block B, (see [4, Section II-
D] for the definition of the KS distance). Function r is defined by
r(.)= 1ifd{o(z)[~(e,)],o ( ' ) ( ~ , ) } > da),r(.)= o otherwise. c(')
is a threshold that can be explicitly determined using statistical tables
associated to the KS limit distribution [4]. Indeed, this is related to
the probability p that the KS distance exceeds &')whereas the two
distributions are the same:
Potential VI favors the labeling of site s with the region number
e, = r if the local distribution at this site is consistent with the
global distribution of the texture features in region r . A reasonable
859
I' c
threshold c(;) may be selected by adjusting the probability p to a
desired level. In practice, large values for p (p > 0.7) enable one
to discriminate textures with close spatial statistics whereas small
values for p (p < 0.3) only discriminate textures with very different
as the determination of the number of regions R, is performed using
a deterministic relaxation algorithm described in the next section.
C. The Relaxation Algorithm
To minimize the global energy function, stochastic relaxation [5]
has not been considered since it is too expensive from a compu-
tational point of view. We have adopted a modified version of the
standard deterministic ICM algorithm [l], which alternately updates
the segmentation map e and the current number of regions.
Let us denote by R the current number of regions. The candjdate
region labels for a given site s are thus labels r = 1, . . . , R. In
the standard ICM algorithm [l], the label that produces the largest
decrease of the energy function is selected to update site s. This
amounts to the minimization of a local energy function that is derived
from the global energy U by computing the energy variation induced
by updating the site s [l], [5]. To cope with the unknown number
of regions, an additional label p corresponding to an "outlier class"
860 IEEE TRANSACTIONS ON IMAGE PROCESSING, VOL. 4, NO. 6, JUNE 1995

II I
II .- 4

II d
Fig. 5. (a) Three-region textured image. (b) Ground truth (the boundaries of the different regions are superimposed on the original image (c) Result of the
unsupervised segmentation procedure ( p = 0.55). (d) Difference between the ground truth and the unsupervised segmentation.

Fig. 6. (a) Aerial image (Calvi area, by courtesy of DRET-Matra mission 1987); (b) result of the unsupervised segmentation procedure (the boundaries
of the different regions are superimposed on the original image) (p = 0.2).

is introduced in the MRF modeling. The outlier label p is assigned all existing regions is too large. More precisely, between the labels of
to site s if the KS distance between the local distribution of texture sites neighboring s and label p . the algorithm selects the label T = i.
features on block B, (centered at s) and the global distributions on that minimizes the following local energy (derived from the global
IEEE TRANSACT1 VS ON IMAGE PROCESSING, VOL. 4, NO. 6. JUNE 1995 86 1

Fig. 7. (a) Aerial image (Roumard area, by courtesy of DRET-INRA mission 1976); (b) result of the unsupervised segmentation prwedL (the boundaries
of the different regions are superimposed on the original image) (JI = 0.65).

Fig. 8. (a) Satellite image of the Alps (France); (b) result of the unsupervised segmentation procedure (the boundaries of the different regions are
superimposed on the original image) (p = 0.25)

energy function U ) : the connected components of the outlier class (i.e., the pixels labeled
p ) are relabeled with new region numbers (Fig. 2). However, label
Forr # p : AU,(r) = ~ [ 1 ~- ( - ert ) ] p is maintained for all small connected components (of size less
tneighborof s than 1% of the image size). These points will not be considered for
m
the computation of global region characteristics, but may be visited
+ ( 2 r [ d ( o ( ' ) ( R ( e= > c(*)]- l} (8)
, r ) ) .o('))(B,)) and relabeled in the next relaxation step. This procedure allows a
2=1
small region to grow until it reaches a size sufficient for providing
F o r r = p : Al',(p) = 11 + 4.
[1- 2S(p - e t ) ] (9) significant statistics. The global statistical distributions associated
CEC with each significant region are updated after the relabeling process.
This relaxation-relabeling process is pursued until stability is obtained
is the parameter of the outlier class (4 > 0 and a high value for
on the label field. The number of iterations required to reach stability
disfavors the creation of new regions). A value of 4 = 2.5 was
depends on the complexity of the scene. However, in practice, the
adopted in all experiments. This local update is applied in turn to the
process has been stopped after three complete relaxation-relabeling
different sites in the image.
procedures with satisfactory results.
For the initialization of the process, an initial value of R = 1 is
used, which corresponds to one single region in the image. This initial
single region is updated using a maximum likelihood estimator of the IV. EXPERIMENTAL
RESULTS
label field, which is nothing but the minimization of the global energy The segmentation technique described in the previous section
without the prior term U2 [7].The current number of regions R is has been applied to different mosaics composed of natural textures
updated once the relaxation on the label field e has converged. At extracted from the Brodatz album, as well as on real-world textured
the end of each relaxation stage (i.e., when the label field is stable), images. Experimental results on images used in other references are
862
presented to make the comparison with other approaches easier. Other
results are reported in 171.
The features related to the joint occurrency matrices were computed
on 16 gray-level images in 5 x 5 local windows, along four directions
with a distance d = 3 (d = 2 yielded similar results). The other
parameters (gray levels, contrast, and directional residuals) were
computed on a 15 x 15 window. Parameter p was set to 1.5, which
corresponds to a good balance between the a priori energy U2 and
the data term U1 . The only parameter that needed to be tuned
was parameter p [defined by (7)], which determines the different
thresholds on the texture features. The results turned out to be
largely insensitive to the adjustment of parameter p (the “optimal”
value could be modified by 10% without noticeable changes in the
segmentation: results obtained for different values of parameter p are
presented in [7]). The CPU time, on a S u n 4 station, reached about
25 mn, for 128 x 128 images.
We first describe the results obtained on various texture mo-
saics, including different textures and object geometry. Fig. 3(a)
presents the first mosaic, which is composed of three different
textures. Fig. 3(b) shows the boundary of the underlying mosaic,
which corresponds to the ground truth. The boundaries of the final
segmentation are superimposed on the original image in Fig. 3(c).
The error between the ground truth and the segmentation obtained
by the unsupervised method is depicted Fig. 3(d). Three regions are
obtained in this case for a value p = 0.75. The boundaries of the
different regions are delineated quite accurately, even if the MRF
prior tends to “smooth” the region boundary in high curvature areas.
Similar qualitative results have been obtained on several other hand-
assembled images corresponding to different mosaic patterns. Fig. 4
reproduces a mosaic pattem close to the one used in the reference
[4] (by Geman et al). A second typical mosaic, borrowed from [lo]
(in which a MRF model is used for representing textures), is shown
in Fig. 5. As can be seen, our result are qualitatively close to the one
reported in [4] and [lo].
The processing of red-world (textured) images is generally more
involved, mainly because the textures appearing in natural images are
usually nonstationary. We present results of segmentation on three
real-world images that have been selected for their representativity
and their difficulty. The images show several textured regions with
various (possibly nonstationary) spatial statistics (Figs. 6-8). Artifacts
or nonstationary areas in textures are difficult to handle and lead
generally to over-segmented images if they are not taken into account
in the model. The approach proposed here, although using stationary
texture models, seems quite robust to these phenomena, as will be
seen. Fig. 6 corresponds to an aerial image showing two cultivated
fields: the boundary between the two regions is delineated quite
accurately, although the upper region is not homogeneous. Fig. 7
presents another aerial image. This kind of image is typical of real-
world applications, in which the textures are nonhomogeneous and
where the number of regions is difficult to determine, even visually.
In this case four significant regions are obtained: in particular, the
forest area has been segmented in two regions, with different spatial
patterns. The two other regions are cultivated fields. Fig. 8 shows
a satellite image of the south-eastem part of France, with similar
difficulties. The area includes a part of the Alps and a part of
the Mediterranean Sea. Even on this hard example (which shows
highly nonstationary textures), a relevant segmentation (Fig. 8(b)) is
obtained. The algorithm enables one to separate quite accurately the
different significant areas: mountain, land, and sea.
V. CONCLUSION
In this correspondence, we have described a texture segmentation
method that can be considered as a first step toward a completely
IEEE TRANSACTIONS ON IMAGE PROCESSING, VOL. 4, NO. 6, JUNE 1995
data-driven segmentation algorithm. The proposed approach does not
require training on the textures to be partitioned. The number of
regions in the image need not to be known either. The method makes
use of local and global spatial statistics within a global statistical
estimation framework relying on Markov random field models. The
method only requires the tuning of one parameter that determines the
thresholds on the different texture features. Several examples have
shown the capability of the model to yield usable segmentations on
hand-assembled images consisting of natural textures as well as on
complex real-world textured images.
REFERENCES
J. Besag, “On the statistical analysis of dirty pictures,” J. Royal Statist.
Soc., vol. 48, sene B, no. 3, pp. 259-302, 1986.
P. Bouthemy and E. Frangois, “Motion segmentation and qualitative
dynamic scene analysis from an image sequence,”Int. J. Comput. Vision,
vol. 10, no. 2, pp. 157-182, 1993.
R. Chellappa and A. K. Jain, Eds., Markov Random Fields: Theory and
Applications. New York: Academic, 1993.
D. Geman, S. Geman, C. Graffigne, and D. Pong, “Boundary detection
by constrainedoptimization,”IEEE Trans. Pattem Anal. Machine Intell.,
vol. 12, no. 7, pp. 609-628, July 1990.
S. Geman and D. Geman, “Stochastic relaxation, Gibbs distributionsand
the Bayesian restoration of images,” IEEE Trans. Pattern Anal. Machine
Intell., vol. PAMI-6, no. 6, pp. 721-741, Nov. 1984.
R. M. Haralick, “Statistical and structural approaches to texture,” Proc.
IEEE, vol. PROC-67, no. 5, pp. 786-804, May 1979.
C. Kervrann and F. Heitz, “A Markov random field model-based
approach to unsupervised texture segmentation using local and global
spatial statistics,” Technical Report 2062, INRIA, Oct. 1993.
A. Khotanzad and J. Y.Chen, “Unsupervised segmentation of textured
images by edge detection in multidimensional features,” IEEE Trans.
Pattem Anal. Machine Intell., vol. 1, no. 4, pp. 414421, 1989.
J. Levy-Vehel, P. Mignot, and J.-P. Berroir, “Multifractals, texture and
image analysis,” in Proc. CVPR’92, Champaign, IL, June 1992, pp.
661-664.
B. S. Manjunath and R. Chellappa, “Unsupervised texture segmentation
using Markov random fields,” IEEE Trans. Pattem Anal. Machine Intell.,
vol. 13, pp. 478482, 1991.
H. H. Nguyen and P. Cohen, “Gibbs random fields, fuzzy clustering, and
the unsupervised segmentation of textured images,” CVGIP: Graphical
Models Image Processing, vol. 55, no. 1, pp. 1-19, Jan. 1993.
J. F. Silverman and D. B. Cooper, “Bayesian clustering for unsupervised
estimation of surface and texture models,” IEEE Trans. Pattem Anal.
Machine Intell., vol. 10, no. 4, pp. 482495, July 1988.
C. S. Won and H. Derin, “Unsupervised segmentation of noisy and
textured images using Markov random fields,” CVGIP: Graphical
Models Image Processing, vol. 54, no. 4, pp. 308-328, July 1992.