The final version of this paper appeared in IEEE Trans.

on Pattern Analysis and Machine

Intelligence, vol. PAMI-12, pp. 211-216, February, 1990.

Texture Segmentation Using Voronoi Polygons

Mihran Tuceryan
Anil K. Jain

Department of Computer Science

Michigan State University
East Lansing, MI 48824-1027
June 26, 1989

Abstract
Texture segmentation is one of the early steps towards identifying surfaces and objects in an image.
Textures considered here are dened in terms of primitives called tokens. In this paper we have developed
a texture segmentation algorithm based on the Voronoi tessellation. The algorithm rst builds the
Voronoi tessellation of the tokens that make up the textured image. It then computes a feature vector
for each Voronoi polygon. These feature vectors are used in a probabilistic relaxation labeling on the
tokens, to identify the interior and the border regions of the textures. The algorithm has successfully
segmented binary images containing textures whose primitives have identical second-order statistics and
a number of gray level texture images.

1 INTRODUCTION
The natural world abounds with textured surfaces. Any realistic vision system that is expected to work
successfully, therefore, must be able to handle such input. The process of identifying regions with similar
texture and separating regions with dierent texture is one of the early steps towards identifying surfaces
and objects. This process is called texture segmentation and is the major focus of this paper.
Texture segmentation has been studied for a long time using various approaches. The most popular
approach performs texture analysis directly upon the gray levels in an image. This includes gray level co-
occurrence matrix (GLCM) [10], autocorrelation function analysis [10], generalized co-occurrence matrices
(GCM) [8], second order spatial averages [9], and two-dimensional ltering in the spatial and frequency
domain [6,5,24,25,20]. Another approach operates at a symbolic level where a textured image is organized
 This research was supported in part by NSF Grant no. IRI 8705256 and Grant no. CDA 8806599.

1
or represented in terms of primitives. Examples of this can be seen in Julesz's work [13,12] and in syntactic
texture analysis. There has also been recent interest in texture analysis approaches based on statistical
modeling such as Markov random elds (MRF) [7,17,16]. Some texture analysis methods, for example, Beck
et al. have examined the role of spatial frequency channels (signal processing level) and perceptual grouping
(symbolic level) in texture segregation [1,2].
Julesz had conjectured [14,15] that texture pairs with identical second-order statistics cannot be preat-
tentively discriminated by humans, but he later gave counterexamples to this conjecture [13]. An important
factor in the preattentive discrimination of such texture pairs appears to be certain shape features of the
texture primitives. These features include closure, linearity, terminations, etc. which are called textons [13].
The important questions that need to be answered in this theory are whether the list of textons is complete
and how to extract textons from images. One approach to this problem is to formulate a general framework
for feature detection so that the detected features capture the texton information. If this is accomplished
then we no longer have to list all the textons. This is the main motivation in this paper. Shape features of
the Voronoi polygons are used to perform texture segmentation and they seem to capture the perceptually
necessary information for this task. Furthermore, the features so extracted depend on local properties of the
image.
The textures that are of interest to us are dened in terms of texture primitives which we call tokens.
These tokens can be either dots or points (simplest case) or they can be more complex primitives such as
arrows and triangles. Primitives can be extracted from the gray level images as a result of some low-level
processing (Section 3.3) or they can be generated synthetically in order to control precisely the various spatial
and statistical properties of the texture [13].
In this paper we propose a method of obtaining shape features from tokens at a symbolic level which will
be useful for texture segmentation. This is accomplished by rst computing the Voronoi tessellation of the
input texture and then extracting the shape properties of the resulting Voronoi polygons. Shape features
of the polygons are represented by their moments of area which are able to capture not only the rst-order
dierences in textures (such as density dierences) but also higher order statistical dierences in textures
(e.g., two textures with the same dot density but dierent orientation tendencies). The Voronoi tessellation
has been used as the underlying geometric representation in other problems such as grouping of dot patterns
[23]. The latter, however, did not involve the computation of textural properties.
Our texture segmentation scheme consists of the following three basic steps: (a) Dene local relationships
between texture primitives or tokens using a graph structure. We use the Delaunay graph (Voronoi diagram)
for this purpose. (b) Identify the degree of \inconsistency" of the edges in the Delaunay graph. The edges
in the interior regions of a homogeneous texture region will be \consistent" and those between two dierent
texture regions will be \inconsistent." The degree of inconsistency of the Delaunay edges are computed
using the shape features of each Voronoi polygon (Section 3.1). (c) The initial labeling in (b) is rened using
probabilistic relaxation labeling, which uses the Gestalt constraint of boundary smoothness. Algorithm 1

2
Input: Gray level texture image.
1. Extract tokens.
2. Construct the Voronoi tessellation.
3. Compute initial BREAK probabilities for Delaunay edges based on moments of area of polygons. (See
Section 3.1).
4. Perform a probabilistic relaxation labeling to identify interior and border tokens. Pi(k) () is the prob-
ability of label  for the ith token at the kth iteration. (See Section 3.2).
5. The nal label on token i is  if Pi(k) () = max P (k) (0 )
0 i
Output: Output labeling identies interior regions and borders of regions.
Algorithm 1: The overall algorithm for the texture segmentation.

shows the steps of the texture segmentation algorithm.

Section 2 describes the Voronoi tessellation and the computation of the textural features from it. Section 3
describes our segmentation algorithm and the probabilistic relaxation labeling scheme. Section 4 gives
experimental results and Section 5 makes some concluding remarks.

2 VORONOI TESSELLATION
Our texture segmentation algorithm uses the Voronoi tessellation to establish local relationships among the
tokens. The features of Voronoi cells are then used to group the tokens belonging to the same texture. In
this section we give a denition of the Voronoi tessellation and the texture features we compute from it.
Let S denote a set of three or more points in the Euclidean plane. Assume that these points are not all
collinear, and that no four points are cocircular. Consider an arbitrary pair of points P and Q belonging
to S. The bisector of the line joining P and Q is the locus of points equidistant from both P and Q and
divides the plane into two halves. The half plane HPQ (HQP ) is the locus of points closer to P (Q) than to Q
(P ). For any given point P , a set of such half planes is obtained for various choices of Q. The intersection
\
HPQ denes a polygonal region consisting of points in the Euclidean plane closer to P than to any
Q2S;Q6=P
other point. Such a region is called the Voronoi polygon [26] associated with the point P . The set of complete
polygons for all points in S is called the Voronoi diagram of S [22]. The Voronoi diagram together with the
incomplete polygons in the convex hull of S dene a Voronoi tessellation of the entire plane. Two points
P; Q 2 S are said to be Voronoi neighbors if the Voronoi polygons of P and Q share a common edge. The

3
 Feature Computation
f1 m00
p
f2 x2 + y 2
f3 tan,1 (y=x)
 1  12
f4 [(m20 ,m02 )2 +41m211 ] 2
[(m20 ,m02 )2 +4m211 ] 2 +m20 +m02
f5 tan,1 (2m11 =(m20 , m02 ))

Table 1: The computation of texture features from moments.

dual representation of the Voronoi tessellation is the Delaunay graph which is obtained by connecting all the
pairs of points which are Voronoi neighbors as dened above.
In the following discussion, rst we consider textures dened in terms of dots and then generalize our
approach to arbitrary tokens. In order to segment the textured image, we compute local features in the dot
pattern and then group dots with similar features to construct uniform texture regions. This also helps in
identifying borders separating such uniform regions. Moments of area of the Voronoi polygons serve as a
useful set of features that re ect both the spatial distribution and shapes of the tokens in the textured image.
The (p + q)th order moments of area of a closed region R with respect to a point or dot with coordinates
(x0 ; y0 ) are dened as [11]: ZZ
mpq = (x , x0 )p (y , y0 )q dx dy (1)
(x;y)2R
where p + q = 0; 1; 2; . . .. These moments are computed with respect to the coordinates of the dot to which
the polygon belongs, so that they are comparable for each polygon across the dot pattern. The details of
ecient computation of these moments for polygonal regions can be found in [27]. The lower order moments
(small values of p and q) have well dened geometric interpretations. For example, m00 is the area of a
polygon, m10 =m00 and m01 =m00 give the displacement of the polygon centroid with respect to (x0 ; y0 ) in
the x and y directions, respectively. The m20 , m11 , and m02 can be used to derive the amount of elongation
of a polygon, and the orientation of its major axis. The higher order moments give even more detailed shape
characteristics of the polygons such as symmetry, etc.
Let (x; y) be the coordinates of the polygon centroid, where x = m10 =m00 and y = m01 =m00 . We have
used the features derived from moments of area shown in Table 1. f2 gives the magnitude of the vector
from the dot to the polygon centroid, f3 gives its direction, f4 gives the overall elongation of the polygon
(f4 = 0 for a circle), and f5 gives the orientation of its major axis. The feature vector for a dot then is
< f1 ; f2; f3 ; f4 ; f5 >.
How do we compute the Voronoi tessellation when the texture primitives are arbitrary shaped tokens
as in Figures 3{5? We have chosen to regard a token as consisting of dots. We compute a generalized

4
Delaunay graph to dene the neighborhood relationship among the tokens and compute the shape feature
for the token in terms of the features of its constituent dots' polygons. Let T be the set of tokens. The
generalized Delaunay graph for the tokens is dened as follows: two tokens ti ; tj 2 T are neighbors or have
an edge between them if there exist two dots pk 2 ti and pl 2 tj such that pk and pl are neighbors in the
Delaunay graph for the dot pattern making up the tokens. The feature vector for the token is computed
based on the features of its constituent dots. The simplest such feature for a token is the average of the
feature for the dots. Note that this is an ad hoc measure, but it appears reasonable for the texture examples
we have used. Let fk be the average of the kth feature. Then we dene the feature vector for the token as
< f 1 ; f 2 ; f 3 ; f 4 ; f 5 >.

3 SEGMENTATION ALGORITHM
In this section we discuss the main steps of the segmentation algorithm. We rst describe computation of
the probability that a Delaunay edge eij , connecting two tokens i and j , is broken (i.e., whether the edge is
inconsistent). Next we describe the probabilistic relaxation labeling to disambiguate and identify the borders
of the regions.

3.1 Initial Break Probabilities on Delaunay Edges

Let eij be the Delaunay edge connecting tokens i and j . Let Ni be the set of neighboring tokens (immediate
neighbors and neighbors of neighbors) for token i. Similarly, dene Nj for token j . If the Delaunay edge
eij lies within a single textured region, then the textural properties of the two sets of neighbors Ni and Nj
should be similar. If eij connects tokens belonging to two dierently textured regions, then the two sets of
neighbors Ni and Nj will have dierent properties. The probability Pij (NB ) that the edge eij is not broken
is dened in terms of the similarity of the features of sets Ni and Nj .
Let Fk = (f1k ; . . . ; f5k ) be the feature vector associated with token k. These are the derived features
discussed in Section 2. We use the Kolmogorov-Smirnov (KS) statistic [19] to compare the set of features
associated with the two neighborhoods Ni and Nj . Let Sk (flk ) be the unbiased estimator of the cumulative
distribution function of the lth feature associated with Nk . If the feature values flk are sorted, then Sk (flk )
is the function giving the fraction of the values to the left of flk . The KS statistic Dijl for comparing the
similarity of the two neighborhoods Ni and Nj based on the lth feature is dened as

Dijl = ,1max jS (x ) , Sj (xl )j:

<xl <1 i l
(2)

Dijl would be close to zero for two neighborhoods which are within a single texture. We use Pij (NB ) =
1 , max
l
fDijl g. It is possible that some of the moments of area (and thus the derived features) will have
similar values even if the two neighbor sets belong to dierent textures. For example, two textures with the
same average dot density, but with dierent orientation trends, will have comparable zeroth- and rst-order

5
moments but dierent second-order moments. We assume the worst case and pick the feature with the
largest dierence in the two cumulative distribution functions to compute Pij (NB ) for the Delaunay edge
eij . A high value of Pij (NB ) means that the two tokens i and j belong to the same texture and thus the
Delaunay edge eij should not be broken. We also dene the probability Pij (B ) that the edge eij is broken
as Pij (B ) = 1 , Pij (NB ).

3.2 Identication of Textured Regions by Probabilistic Relaxation Labeling

The probabilities Pij (NB ) computed in the previous section are a good starting point for the region seg-
mentation, but they rely only on local information and do not enforce Gestalt constraints such as border
smoothness. To smooth local irregularities and enforce border smoothness, we formulate a probabilistic
relaxation labeling scheme.
Probabilistic relaxation labeling is a technique of parallel constraint propagation for obtaining locally
consistent labels for a set of objects [21,28]. It consists of a set of objects T = fti ; i = 1; 2; . . . ; ng to be
labeled and a set  = fi ; i = 1; 2; . . . ; mg of labels. Let Pi () be the probability that the label  is \correct"
X
for the ith object, such that Pi () = 1.

The probability Pi(k+1) () of label  on object i at iteration (k + 1) is obtained by using the following
rule: (k) (k)
Pi(k+1) () = XPi (k())[1 + qi (k())]0 ; (3)
Pi ()[1 + qi ( )]
0
where  
X
qi k () = wij max
( )

0 (k) 0
0 frij (;  )Pj ( )g + min

0 (k) 0
0 frij (;  )Pj ( )g : (4)
j
In Equation 4, rij (; 0 ) is the compatibility of the labels  and 0 on objects i and j and wij is the coecient
which determines the contribution of the j th object towards the change in probability for the ith object. The
rij values incorporate the border smoothness constraint. The quantity qi(k) () is in the range [-1,1], and
it determines whether Pi(k+1) () is increased, decreased, or remains unchanged after the probabilities are
normalized. The iteration continues until either (i) all the probabilities in Equation (3) converge, or (ii) the
number of iterations exceeds a prespecied limit.
We now look at the specic formulation of the relaxation labeling for our segmentation problem.
(a) The set of objects to be labeled are the n tokens in the texture image, T = fti ; i = 1; 2; . . . ; ng.
(b) The set of labels  = fI; E; N; W; S g, indicate whether the token is an interior token (I ), or it lies on
the border of a region with direction 0 degrees (E ), 90 degrees (N ), etc. The sense of the direction
is chosen according to the right hand rule; that is, when facing in the direction of the border label,
the interior of the region lies on the right hand side. In this algorithm we have decided to use four
quantized directions for the border labels. One can increase the resolution by using a larger number
of border directions.

6
(c) The initial label probabilities are computed from the initial Pij (NB ) computed before. For example,
X
Pi(0) (I ) = jN1 j Pij (NB ) (5)
i j 2N i

This is based on the expectation that in the interior of a homogeneous textured region most Delaunay
edges will have high values of Pij (NB ). The initial probability for a border label bk 2 fE; N; W; S g is
computed by
( PNB + PB )
Pi(0) (bk ) = WNB 2 WB (6)
where
X 2( 2 , )
PNB = Pij (NB )  (7)
j
X 2( 2 , )
WNB =  (8)
j
X
PB = Pij (B ) 2 (9)
j
X
WB = 2 (10)
j
In the above expressions, the terms PNB =WNB and PB =WB are weighted averages of the Pij (NB )
and Pij (B ) probabilities, respectively, around the token i. The weights depend on the angle  that the
Delaunay edge eij makes with the border label bk (see Figure 1(a)). Edges are expected to be linked
along the border label direction (thus the weight 2( 2 , )= for the NB label) and they are expected
to be broken away from the border label direction (and hence the weight 2= for the B label). The
angle  can take on values in the range [0; 2 ]. The initial probabilities for those dots whose Voronoi
polygons are incomplete are computed dierently, because in these cases we know a priori that the
dot is not an interior dot. This computation takes advantage of the fact that the interior regions lie
on the non-empty side of the border. In this case, we assign Pi(0) (I ) = 0, and proceed with the above
analysis for the probabilities of border labels.
(d) The Gestalt constraint of border smoothness is enforced through the proper use of compatibility coef-
cients. These compatibilities between pairs of neighboring tokens are computed based upon the local
evidence for the token being in the interior or lying on the border and the enforcement of local border
smoothness to obtain more perceptually meaningful segments. In the following formulas, Pi (bk ) is the
probability of token i having border label bk 2 fE; N; W; S g, ij is the angle formed by the edge from
token i to token j with the x-axis, and k is the angle of the border label bk with the x-axis (see Fig-
ure 1(b)). The compatibilities rij (; 0 ) between two tokens i and j with labels ; 0 2 fI; E; N; W; S g
are computed as follows:
rij (I; I ) = Pi (I ) +2 Pj (I ) (11)

7
 bk bk
. .
6 ... .... 6k ... ....
.... .... ij
... .... bl
... ....
.
 .

A
K
  j A l  j
 A 
 A
i i
(a) (b)

Figure 1: (a) The denitions of angles in the interior-border compatibilities. (b) The denitions of angles in
the border-border compatibilities.

 1 + sin( ,  )  P (I ) + P (b )
rij (I; bk ) = ij k i j k (12)
2 2
  P (I ) + P (b )
rij (bk ; I ) = 1 , sin(2ij , k ) j 2 i k (13)
 
rij (bk ; bl ) = Pi (bk ) +2 Pj (bl ) Sangle (14)
where   
Sangle = 21 1 + cos 2(k , ij ) +
2
cos 2(l , ij ) cos( ,  )
k l (15)

In these expressions, Pi and Pj provide the local evidence for the particular labels whose compatibilities
are computed. The sin(:) terms in Equations (12) and (13) enforce the right hand rule for the border
direction. That is, the support provided by the sin(:) term is positive if the interior labeled dot is on the
right hand side of the border direction. Sangle in Equation (14) enforces two conditions: (i) When the
border tokens are along the contour of a single region, they must be aligned and pointing in the same
direction (k , ij = l , ij = k , l = 0). (ii) When the border tokens are along the contours of two
separate regions, they should be pointing in opposite directions (e.g., ij = 0, k = =2, l = ,=2).
This is done in order to keep the interior of each region on the right side of each border label. The
space between two such border tokens is the crack between two texture regions.
This relaxation labeling algorithm is run on the tokens to obtain the nal labeling of the interior regions
and the borders of the texture. In running the relaxation labeling algorithm, we update our compatibility
coecients after a certain number of iterations (5 iterations in this case). This updating re ects the eect of

8
changing probabilities of the labels on the compatibilities of neighboring tokens. The relaxation algorithm
is run for 100 iterations in total. In most instances the label probabilities converge to their nal values in
fewer iterations. However, for certain ambiguous cases, it takes longer for the values to settle.

3.3 Gray Level Texture Images

In order to apply our texture segmentation algorithm to gray level images, we need to rst extract tokens
from images. Some researchers have developed complex methods for extracting such texture elements from
gray level images [3,25]. In this paper, we have used a simple algorithm to extract tokens from texture
images to demonstrate that our texture segmentation algorithm can be applied not only to synthetic dot
patterns but to gray level images as well.
We extract tokens from a gray level image as follows:
1. Apply Laplacian of a Gaussian (LoG or r2 G) lter to the image. For computational eciency, we
approximated the r2 G lter with a dierence of Gaussians (DoG) lter. The size of the DoG lter
is determined by the sizes of the two Gaussian lters. We used 1 = 1 for the rst Gaussian and
2 = 1:61 for the second one. According to Marr, this is the ratio at which a DoG lter best
approximates the corresponding r2 G lter [18, page 63].
2. Select those pixels that lie on a local peak of the ltered image. A pixel in the ltered image is said
to be on a local peak if its magnitude is larger than six or more of its eight nearest neighbors. This
denition allows us to identify those pixels that lie along a ridge. This results in a binary image (e.g.
Figure 4(c)).
3. Perform a connected component analysis on the binary image using eight nearest neighbors. Each
connected component denes a texture primitive (token). We have selected a fairly simple method of
token extraction. Several other methods of token extraction have recently been proposed [3,25].
After the texture tokens are identied using this method, we apply the algorithm described in Sections 3.1
and 3.2 to segment the texture image.

4 EXPERIMENTAL RESULTS
We have tested our segmentation algorithm on three classes of texture patterns. The rst class consists of
textures which are made of dots, an example of which is seen in Figure 2(a). These textures dier in the
regularity of the dot placement, the direction of the dot density, and the average dot density. In some of
the regular textures we have added small amounts of perturbations to the positions of the dots. The second
class of textures have primitives made up of dots. Figure 3(a) gives a popular example of such textures seen
in the literature [13]. In this example, the tokens which are used to construct the textures have identical

9
second-order statistics. The third class of texture patterns were taken from the texture album of Brodatz
[4]. Several gray level images of textures were put together in pairs to test our segmentation algorithm.
Figures 4(a){5(a) are examples of such texture pairs. These are 128x256 images (each Brodatz texture is
128x128) and they have a range of 256 gray levels.
Our segmentation algorithm can successfully segment the textured regions of the rst class (Figure 2(a)).
Figure 2(b) shows the probabilities Pij (NB ) for the Delaunay edges for the textures in Figure 2(a). The
probabilities are shown as gray levels such that the lighter edges have higher probabilities of being broken.
As can be seen from Figures 2(a) and 2(b), the initial estimates of Pij(0) (NB ) are quite useful in identifying
the interior regions (where most edges are not broken) and the borders between textured regions (where
most Delaunay edges are weak). The nal segmentation after running the relaxation labeling is shown in
Figure 2(c). In the segmentation results the tokens are represented as dots at their centroids for graphical
representation. Note that the borders in Figure 2(c) are thick and fuzzy due to the local nonuniformity of
the shapes of the Voronoi polygons going from one region to the next.
Our algorithm also successfully segments the \corner-closure" texture pattern (Figure 3(a)). In Fig-
ure 3(b), we are able to successfully label the interiors of the two textures and identify the correct location
of the border. There are some textures with identical second-order statistics that our algorithm is unable to
segment, but these textures are not preattentively discriminable by humans, either. We have also run the
segmentation algorithm on patterns that consist of textures made up of L's and crosses, L's and T's, etc. [1].
In each of these cases, the algorithm nds the borders and the interiors of the textures properly.
Finally, the algorithm successfully segments most of the texture pairs we used from the Brodatz texture
album [4]. Figures 4(a){5(a) give the input gray level images. Figures 4(b){5(b) give the result of convolving
these images with the DoG lter (1 = 1, 2 = 1:61 ). Figures 4(c){5(c) give the peaks that are detected from
the convolved images. These are the tokens used to segment the textures. The results of the segmentation
is shown in Figures 4(d){5(d). In these segmentation results also the tokens are represented as dots at their
centroids for graphical representation.
The segmentation results of Figures 4 and 5 detect the borders and the interior regions of the texture
pairs. There are some tokens in the interior regions of some of the textures (Figure 5) that are labeled as
borders. There are two reasons for this: (a) Our comparison of statistical similarity of features is local and
thus it is detecting the dierences in these local regions, (b) The tokens actually have dierent local properties
(e.g. in Figure 5(c), the wood grain texture has long vertical tokens intermixed with short and horizontal
tokens). In some texture pairs (e.g., D9 grass lawn and D17 herringbone weave), complete segmentation is
not achieved but a large portion of the border between the two textures is detected. We feel that better
token extraction, increasing the size of the neighborhoods around each token, and some postprocessing will
improve the performance of our segmentation algorithm.

10
5 CONCLUSION
In this paper we have developed a texture segmentation algorithm based on the Voronoi tessellation. The
algorithm rst builds the Voronoi tessellation of the tokens that make up the textured image, computes a
set of features from the Voronoi diagrams, and nally, it performs a probabilistic relaxation labeling on the
tokens, to identify the interior and the border regions of the textures present in the image.
The results of the segmentation algorithm show that the shapes of the Voronoi polygons provide a
powerful set of features that re ect certain textural properties in images. Particularly, the results on the
\corner-closure" type texture patterns of Figure 3 show that the important aspects of the textural properties
have been captured by this representation.
An important advantage of our segmentation algorithm over other texture segmentation algorithms is
that we do not have to specify the number of texture regions (segments) a priori. The number of regions
is automatically found by identifying contiguous interior regions. The moment-based features are able
to capture orientation sensitive texture properties, as well as more complicated textural properties such
as textons proposed by Julesz [13] (see, for example, Figure 3(a)). Our segmentation algorithm tries to
incorporate Gestalt perceptual criteria such as border smoothness. In many of the texture patterns, these
Gestalt criteria improve the segmentation results as seen in Figure 2.
Future research directions range from improvements in the current algorithm to its integration with other
intermediate level visual processing modules. We need to improve the method used in comparing the textural
properties of the two neighborhoods of the Delaunay edge (Section 3.1). The current algorithm is sensitive
to small perturbations in the locations of the tokens if the token placement is very regular. The fact that in
the regular regions the variance of the features is almost zero makes the statistical comparison very sensitive
to small perturbations in such areas. This results in border labels being assigned to interior tokens because
slight perturbations in the location of the tokens cause the feature distributions to be seen as dierent.
The localization of the actual boundary has not been addressed in this paper. Some post processing is
also necessary to thin the boundary and smooth the texture regions. This step may also improve the results
for regular regions with very small perturbations discussed above.
In the case of the gray level texture images, we used a very simple method of token extraction. This
process can be improved further which will result in better segmentation. The use of multiscale ltering in
the token extraction process is one possibility.
A major extension of this work would be in the integration of the segmentation process with other
visual processing modules. One immediate area is the shape from texture processing where the boundaries
obtained from the texture segmentation may be used in localizing the process of surface geometry extraction.
A feedback from the shape from texture algorithms may in turn provide useful information about possible
borders between textured regions.

11
ACKNOWLEDGEMENTS
The implementation of the algorithm described in Sections 3.1 and 3.2 was partly done by Mrs. Ywhyng
Lee Homan. Miss Sally Howden has also helped in parts of the implementation.

References
[1] J. Beck, K. Prazdny, and A. Rosenfeld. A theory of textural segmentation. In Jacob Beck, Barbara
Hope, and Azriel Rosenfeld, editors, Human and Machine Vision, pages 1{38, Academic Press, New
York, 1983.
[2] J. Beck, A. Sutter, and R. Ivry. Spatial frequency channels and perceptual grouping in texture segre-
gation. Computer Vision, Graphics, and Image Processing, 37:299{325, 1987.
[3] D. Blostein and N. Ahuja. Representation and three-dimensional interpretation of image texture: an
integrated approach. In Proc. of First International Conference on Computer Vision, pages 444{449,
1987.
[4] P. Brodatz. Textures: a photographic album for artists and designers. Dover Publications, Inc., New
York, 1966.
[5] M. Clark and A. C. Bovik. Texture segmentation using Gabor modulation/demodulation. Pattern
Recognition Letters, 6:261{267, Sept. 1987.
[6] J. M. Coggins and A. K. Jain. A spatial ltering approach to texture analysis. Pattern Recognition
Letters, 3:195{203, 1985.
[7] G. C. Cross and A. K. Jain. Markov random eld texture models. IEEE Trans. on Pattern Analysis
and Machine Intelligence, PAMI-5:25{39, 1983.
[8] L. S. Davis, M. Clearman, and J. K. Aggarwal. An empirical evaluation of generalized cooccurrence
matrices. IEEE Trans. on Pattern Analysis and Machine Intelligence, PAMI-3:214{221, 1981.
[9] A. Gagalowicz. Blind texture segmentation. In Proc. of the 9th International Conference on Pattern
Recognition, pages 46{50, Rome, Italy, November 1988.
[10] R. M. Haralick. Statistical and structural approaches to texture. Proceedings of the IEEE, 67:786{804,
1979.
[11] M. K. Hu. Visual pattern recognition by moment invariants. IRE Trans. on Information Theory,
IT-8:179{187, 1962.
[12] B. Julesz. Texton gradients: the texton theory revisited. Biological Cybernetics, 54:245{251, 1986.

12
[13] B. Julesz. Textons, the elements of texture perception, and their interactions. Nature, 290:91{97, 1981.
[14] B. Julesz. Visual pattern discrimination. IRE Trans. on Information Theory, IT-8:84{92, 1962.
[15] B. Julesz, E. N. Gilbert, L. A. Shepp, and H. L. Frisch. Inability of humans to discriminate between
visual textures that agree in second-order statistics - revisited. Perception, 2:391{405, 1973.
[16] R. L. Kashyap and R. Chellappa. Estimation and choice of neighbors in spatial interaction models of
images. IEEE Trans. on Information Theory, 29:60{72, 1983.
[17] R. L. Kashyap, R. Chellappa, and N. Ahuja. Decision rules for the choice of neighbors in random eld
models of images. Computer Graphics and Image Processing, 15:301{318, 1981.
[18] D. Marr. Vision. W. H. Freeman and Company, San Francisco, CA, 1982.
[19] W. H. Press, B. P. Flannery, S. A. Teukolsky, and W. T. Vetterling. Numerical Recipes in C. Cambridge
University Press, New York, 1988.
[20] T. C. Rearick. A texture analysis algorithm inspired by a theory of preattentive vision. In Proc. of
IEEE Conf. on Computer Vision and Pattern Recognition, pages 312{317, San Francisco, CA, 1985.
[21] A. Rosenfeld, R. Hummel, and S. Zucker. Scene labeling by relaxation operations. IEEE Trans. on
Systems, Man, and Cybernetics, SMC-6:420{433, 1976.
[22] M. I. Shamos and D. Hoey. Closest-point problems. In Proc. 16th Annual Symp. on Foundations of
Computer Science, pages 131{162, 1975.
[23] M. Tuceryan and N. Ahuja. Extraction of early perceptual structure in dot patterns: integrating region,
boundary, and component gestalt. Computer Vision, Graphics, and Image Processing, 1989. To appear
in December.
[24] M. R. Turner. Texture discrimination by Gabor functions. Biological Cybernetics, 55:71{82, 1986.
[25] H. Voorhees and T. Poggio. Detecting textons and texture boundaries in natural images. In Proc. of
First International Conference on Computer Vision, pages 250{258, London, 1987.
[26] G. Voronoi. Nouvelles applications des parametres continus a la theorie des formes quadratiques.
deuxieme memoire: recherches sur les parallelloedres primitifs. J. Reine Angew. Math., 134:198{287,
1908.
[27] H. B. Wilson Jr. and D. S. Farrior. Computation of geometrical and inertial properties for general areas
and volumes of revolution. Computer Aided Design, 8(4):257{263, 1976.
[28] S. W. Zucker and R. A. Hummel. On the foundations of relaxation labeling processes. IEEE Trans. on
Pattern Analysis and Machine Intelligence, PAMI-5:267{287, 1983.

13
.......................... ..... .... ....
....... ...
.. .........
.
.
.
..
....
......
......
......
......
......
......
......
......
......
...
......
...
.. . ..
. ....
.
..
. .
.
.
.
.
...
. ..
. ...
.. .
.
..
.
.
.................
. .
..
...
..
. ..
.. ..
.. .
......................... .. . .. . . . .
.
...
. ..
.. ..
.. ..
.. ..
.. ..
.. ..
.. ..
.. ..
.. ..
.. ..
.....
.... . .
...
....
....
... ...
......
.. .
. . .
.......
.. .
.
..
.
...
......
.......................... . . . . .
.........
......
.. .
....... . . ..
.....
.
....
.. ..
.. ..
.. ..
.. ..
.. ..
.. ..
.. ..
.. ..
....
...
.......
... . .. .
.. . . .
.. .. ..
.
....
. . . .
.
.
.....
......
....
....
....
....
....
...
....
......
.. ......
.. . .
............. .
..... ... .
.. ....
..
.
.....
....
....
....
....
....
...
..
..
.......
..... ... .
..... .
......... ... . .
...
.. ... . ..
.
...
....
....
....
....
....
...
..
..
....
...
.........
.. .
. .............
.....
.... . ..... .
.. .
...
... ..........
... . . . . .... .
..
. . . . .. . .
...
.....
....
....
....
...
..
..
..
....
....
...
....
....
........ . .. .
..... .
... .
........... .. ..
...... .
.
....
.. ..
.. ..
.. ..
....
...
.....
.....
...
...
....
....
...
. .......
..... ........ .
........ ........
. .
....
......... . . ... . . .
...
...
.......
.....
.....
...
....
....
....
....
...
......
..
....
......................
. ..
.
.... ..
.
..
...
...
..
...
...
...
..
...
....
....
....
....
...
..
. . ... .. .....
. .. ...
...
.. .. ... ...... .
.
....
....
...
.....
....
...
....
....
....
....
....
....
...
.
.
... .
.....
. . .
........
. ..
.... ...... .
.
....
.
.
..
..
...
.....
....
...
....
....
....
....
....
....
...
..
..
. .
. .
. .
. .
. .
. .
. .
. .
. .
. .
. .
. .
.
.
...
....
...
..
...
....
....
....
....
....
....
....
...
..
..
. .
. .
. .
. .
. .
. .
. .
. .
. .
. .
. .
. .
.
.
..
..
..
..
...
....
....
....
....
....
....
....
....
...
..
..
. .
. .
. .
. .
. .
. .
. .
. .
. .
. .
. .
. .
.
.
...
....
....
....
....
....
....
....
....
....
...
..
..
..
..
. .
. .
. .
. .
. .
. .
. .
. .
. .
. .
. .
. .
.
.
...
....
....
....
....
....
....
....
....
....
....
...
..
..
. .
. .
. .
. .
. .
. .
. .
. .
. .
. .
. .
. .
.
.
...
....
....
....
....
....
....
....
....
....
....
....
...
. .
. .
. .
. .
. .
. .
. .
. .
. .
. .
. .
. .
.
.
...
....
....
....
....
....
....
....
....
....
....
....
...
. .
. .
. .
. .
. .
. .
. .
. .
. .
. .
. .
. .
.
.
...
....
....
....
....
....
....
....
....
....
....
....
...
. .
. .
. .
. .
. .
. .
. .
. .
. .
. .
. .
. .
.
.
...
....
....
....
....
....
....
....
....
....
....
....
...
. .
. .
. .
. .
. .
. .
. .
. .
. .
. .
. .
. .
.
.
...
....
....
....
....
....
....
....
....
....
....
....
...
. .
. .
. .
. .
. .
. .
. .
. .
. .
. .
. .
. .
.
.
...
....
....
....
....
....
....
....
....
....
....
....
...
. .
. .
. .
. .
. .
. .
. .
. .
. .
. .
. .
. .
.
.......................... . . . . . . .
.......................... . . . . . . . . . . . . . . . . .
(a) (b)

............................. . ... ....

.
..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
. .. ....
...
....
.
.
...
. .
..
......
.....
. .
...
...
..
..
.
..
..
....
...................... . ...... . . ...
. ....... . . ..
.. ...... . . .
... ..
.
..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
. .. . . .. . . .. . . .. .. .
. .
..
...................... . ............... . .. .
.......... .
.. .. ..
.
.
..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
. .. .
. ..... ..
. .
...
.. . . . . ........
..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
.. . .
. ..... .. . . ..
.............. .
.
..... .
. .
..
. ..
.
..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
. ..... . ..
........ . . . . .
. ...
.. .. ...
..
. ... . .
..
..
...
..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
... .....
.. .
........... ..
... ..
.
.
..
.. .
...
. ..
. .
..
...
..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
... ...
.. ....
...
...
. ..
..... ..
.
.
......
.. ..
.
.
..
. ...
....
........
..
.
..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
. ........ . .......... ....
...................... . .. .. . .
..
. .
..
. .... .....
. .
...................... ..
. .
. .. .
........ . ..
.. ...... ...
.
..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
. . . .. ....... . . .. .
.
..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
.. .... . . .
.. . .
.
..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
. . . . . . . . . . .
.
..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
. .
. .
. .
. .
. .
. .
. .
. .
. .
. .
.
.
..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
. .
. .
. .
. .
. .
. .
. .
. .
. .
. .
.
.
..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
. .
. .
. .
. .
. .
. .
. .
. .
. .
. .
.
.
..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
. .
. .
. .
. .
. .
. .
. .
. .
. .
. .
.
.
..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
. .
. .
. .
. .
. .
. .
. .
. .
. .
. .
.
.
..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
. .
. .
. .
. .
. .
. .
. .
. .
. .
. .
.
.
..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
. .
. .
. .
. .
. .
. .
. .
. .
. .
. .
.
.
..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
. .
. .
. .
. .
. .
. .
. .
. .
. .
. .
.
.
..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
. .
. .
. .
. .
. .
. .
. .
. .
. .
. .
.
.
..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
. .
. . . . . . . . . .
....................... ..........
. . . . . . . . .

(c)
Figure 2: (a) A texture with three regions. Two regions consist of regularly placed dots with dierent
densities. In the third region the dots are randomly placed. One of the regularly placed regions has a
directionally sensitive density. (b) Pij(0) (NB ) assigned to each Delaunay edge for the tokens in (a). The
probabilities are shown as gray levels such that the lighter edges have higher probabilities of being broken.
(c) Segmentation showing the interior regions (shown as dots) and the borders (shown as arrows in four
directions) for the texture in (a).

14
 . . . . . .
. . . . . .
. . . . . . . . .
. . . . . . . . .
. . . . . . . . . .
. . . . . . . . . .
. . . . . . . . .
. . . . . . .
. . . . . . . .
. . . . . . . .
. . . . . . .
. . . . . . .

(a) (b)

Figure 3: (a) An example texture pair with identical second-order statistics that can be preattentively
discriminated. (b) The resulting segmentation produced by the algorithm. The dots are the place holders
for the token and represent interior regions.

15
 (a) (b)

. ... . ... ..... . . .. . ... .... . .

. . . . ....... . ... ... ...... . .. ..
. ... ..... . .... .... .
. . . . . . .
.. . . ...
. .
. . .. . ... . .... ...
. ....
.... . .. . . .
. ...
. ...... . ..........
. . . .
...... ...
. .. . . ............
.
..... . ... ... . . . . .........
.
. . .. . .
.
..... .. . . ............ ..
.
. . . . .... ................ ..... ...
... . . ..... . . . . . . . ... .
. .. .. .
..... .. .... ............... .... . .
..
. .... ....
..
. .. .. .
. .. .
. . . . .....
. . ......
.. . . .. . .. .. .. . .
. .. ....
.
. . . . . .
.
.
.....

(c) (d)

Figure 4: (a) Texture pair D3 (reptile skin) and D17 (herringbone weave) from Brodatz texture book. (b)
Image in (a) convolved with a DoG lter with 1 = 1 and 2 = 1:6. (c) Detected peaks in the ltered image.
The connected components form the tokens. (d) Resulting segmentation. Each token is displayed as a dot
at its centroid.

16
 (a) (b)

... . . . ... ....... ..... ... .

.
. ..
.. .. ... ... .
.
. ..
. . . . . . ...
... ....
.. .
...... .. .. .. . ... .
.
.
. .. ....
.
. .
.
. .. .
. ...
.
. .. .
.. . .. ...
. .. ..
. .. .
.. ..
. . ..
.
.. .
. . ... . . . .. .. .
. . . .
.... .... . ... ... . ..
.. ... . . . .. .. .. .
. . . . . . .
. . . .
............. . ..
.. ... .
. .... ..
.... ... .. .
.... ... .. ... .. .
.. ...
... . . .
......... . .
... . . .... ...
... .
.
....
.
.
..
.
. .. ... . ..
.
.... ..
. . . . . .
.
.
.. ... ....... ... .... . .
. .... ..... ... ...
. . ... . .. . . . .... . .
. ......... .. . .. .. .. ......
... .. ... .. . .
. ..... ............ . ... . ...
...
. .....
. ..
..... .. .
. . .. .
. .. .
..

(c) (d)

Figure 5: (a) Texture pair D55 (straw matting) and D68 (wood grain) from the Brodatz texture book. (b)
Image in (a) convolved with a DoG lter with 1 = 1 and 2 = 1:6. (c) Detected peaks in the ltered image.
(d) Resulting segmentation. Each token is displayed as a dot at its centroid.

17