Pattern Recognition Letters 23 (2002) 11791190

www.elsevier.com/locate/patrec

Motif analysis of noisy regular textures

Gyuhwan Oh a, Seungyong Lee
a

b,*

Department of Computer Science, Korea Advanced Institute of Science and Technology (KAIST), Taejon 305-701, South Korea
b
Department of Computer Science and Engineering, Pohang University of Science and Technology (POSTECH),
Pohang 790-784, South Korea
Received 25 July 2001

Abstract
A regular texture contains a periodic arrangement of a determinable pattern. Motifs of a regular texture are the
minimal parallelogrammic sub-regions that can reproduce the texture by tiling. A motif extracted from a regular texture
can be used for texture reproduction, texture recognition, and texture compression. However, it is not simple to accurately capture a motif of a regular texture, as the motif may be non-rectangular or arranged in an arbitrary direction.
Moreover, when the texture contains additive noises, such as are created by the scanning process, motif analysis becomes more dicult. In this paper, we present an algorithm to obtain a motif of a noisy regular texture in Omn log mn
time, where m  n is the size of the texture. The analysis examples demonstrate that the proposed algorithm can robustly derive motifs of noisy regular textures.
2002 Elsevier Science B.V. All rights reserved.
Keywords: Regular texture; Motif analysis; Translation symmetry; Distance matching function

1. Introduction
A pattern is a general notion that represents an
organic structure. A regular texture is a rectangular image which contains the repetition of a visually recognizable unit pattern. Due to structural
regularity, a regular texture exhibits an ornamental eect. In commercial areas, such as apparel and
interior design, regular textures have been widely
used as a fundamental resource to achieve decorative eects (Porter, 1975; Stevens, 1991; Washburn and Crowe, 1992). In computer graphics,

Corresponding author. Tel.: +82-54-279-2245; fax: +82-54279-5699.

E-mail address: [email protected] (S. Lee).

they may be used for texture mapping, which is

a low-cost technique to add a realistic appearance
to the surface of an object.
A regular texture can be produced from scratch
by designing a unit pattern and tiling the pattern in
a rectangular region. In this case, however, designing a unit pattern is time-consuming, as it
requires not only the original idea but highly
developed skills as well. Yet, a variety of sources
exist for obtaining regular textures, such as pattern
books (Stevens, 1991; Grafton, 1992; Stegenga,
1992). We can derive an image of a regular texture
by digitizing with a scanner or a digital camera,
and a new texture can be generated by manipulating the scanned image.
Motifs of a regular texture are the minimal subregions of the texture which can regenerate the

0167-8655/02/$ - see front matter
2002 Elsevier Science B.V. All rights reserved.
PII: S 0 1 6 7 - 8 6 5 5 ( 0 2 ) 0 0 0 6 4 - 8

1180

G. Oh, S. Lee / Pattern Recognition Letters 23 (2002) 11791190

texture by tiling. To process a scanned regular

texture more eectively, it is necessary to extract a
motif from the texture. For example, when we
repeat a scanned regular texture to generate a large
texture, we are confronted with the problem of
visual discontinuity along the boundary (see Figs.
1(a) and (b)). This problem can be avoided by
tiling a motif of the texture (see Figs. 1(c) and (d)).
In addition, by modifying a motif of a regular
texture and tiling the result, we can easily obtain a
new regular texture that is similar to the original.
However, it is not simple to accurately capture the
motif of a regular texture. Extra attention is also
required when the motif is either non-rectangular
or arranged in an arbitrary direction (see Fig. 2).
A digitized regular texture usually contains additive noises due to the scanning process. Although
the noises are almost imperceptible to the human
eye, they may destroy the regular structure of the
texture in terms of pixel values. That is, the pixel
values may not exactly match each other between
the repetitions of a visually recognizable motif.
Consequently, when a digitized texture appears to
be regular, it may not be truly regular when we
compare pixel values. Thus, it is more dicult to
analyze the motif of a digitized regular texture.

In addition to texture production, motif extraction can be used for compression and recognition of regular textures. To save space for storing
a regular texture, we can remove redundant information in the texture by retaining only a motif
of the texture. When we compare two regular
textures, we do not need to compare all pixel information in the textures. Motifs extracted from
the textures contain sucient information for determining the equality of the textures.
In this paper, we propose an algorithm to identify a motif of a regular texture in Omn log mn
time, where m  n is the size of the texture. The algorithm can properly handle a noisy regular texture
which preserves a visually regular structure, although its pixel values are not exactly repeated. The
analysis examples demonstrate that the algorithm
robustly derives the motifs of noisy regular textures.
Symmetry group theory provides a mathematical foundation to dene a regular texture and its
motif (Armstrong, 1988; Baglivo and Graver, 1983;
Conners and Harlow, 1980; Martin, 1982). To
analyze the motif of a regular texture, we rst
adopt the properties of a motif from the theory. To
apply the properties to a noisy regular texture, we
use the distance matching function (Oh et al.,

Fig. 1. Repetition of a regular texture: (a) texture; (b) simple repetition; (c) motif; (d) repetition of the motif.

Fig. 2. Regular textures with: (a) non-rectangular; (b) slanted; (c) non-rectangular and slanted motifs.

G. Oh, S. Lee / Pattern Recognition Letters 23 (2002) 11791190

1999), which can approximate the regular structure

of a noisy texture. Several techniques, such as
Delaunay triangulation and least-square approximation, are then used to extract a motif by rening
possible errors in the approximated regular structure.
The remainder of this paper is organized as
follows. Sections 2 and 3 review related work and
symmetry group theory, respectively. In Section 4,
we present the basic approach of the proposed
technique for motif analysis of a noisy regular
texture. Sections 5 and 6 explain the details of two
steps in the technique: lattice approximation and
basis vector approximation. Section 7 gives several
analysis examples. We conclude this paper in Section 8.

2. Related work
A texture can be viewed as a two-dimensional
string that contains characters from a nite alphabet. Ecient algorithms (Galil and Park, 1992,
1996), which are based on string matching techniques, have been proposed to determine the motif
of a regular texture. For a texture of size m  n, the
algorithms can determine a motif in Omn time.
However, the algorithms are complicated and
dicult to implement. Furthermore, no mechanism was provided to extract a motif from a noisy
regular texture, because the algorithms are based
on the equality test of two alphabets.
Several algorithms based on co-occurrence
matrices are commonly used to detect the periodicity of a noisy texture (Conners and Harlow,
1980; Parkkinen and Selkainaho, 1990; Zucker
and Terzopoulos, 1980). The algorithms rst
compute co-occurrence matrices for certain displacement vectors and then calculate measures
such as v2 statistics, j statistics, and inertias from
those matrices. Computing co-occurrence matrices
for all possible displacement vectors requires
Om2 n2 time, where m  n is the size of the texture. However, the algorithms primarily focus on
detecting one-dimensional periodicity and did not
address how to extract a motif from the texture,
which is related to periodicity in two independent
directions.

1181

Oh et al. (1999) proposed the distance matching

function to compute the inertias without constructing co-occurrence matrices, which was made
possible due to their new interpretation of an inertia. For a texture with the size m  n, the inertias
of all co-occurrence matrices can be obtained in
Omn log mn time by simultaneously evaluating
the distance matching function for all displacement vectors. We use the function to obtain the
approximate lattice of a noisy regular texture in
Section 5.

3. Preliminary
In this section, we briey review symmetry
group theory, which provides mathematical denitions of a regular texture and its motif (Armstrong, 1988; Baglivo and Graver, 1983; Conners
and Harlow, 1980; Martin, 1982).
A texture can be considered as a bivariate
function g. The function value gx; y represents
the gray level at position x; y 2 Z2 , where Z is the
set of integers. We assume that g is axis-aligned in
the two-dimensional plane and 0; 0 is the lower
left corner of the texture.
Let R represent the set of real numbers. A
translation sv by a vector v vx ; vy is a function,
sv : R2 ! R2 , such that sv p p v px vx ;
py vy for a point p px ; py 2 R2 . When a
translation sv is applied to a texture g, it translates
every pixel in g by v. That is, if g is of size m  n
and we let g0 sv g, then g0 is a texture of the
same size as g, and g0 x; y gx vx ; y vy for
x vx ; vx 1; . . . ; vx m 1 and y vy ; vy 1;
. . . ; vy n 1.
The composition sv su of two translations su and
sv is the translation suv by the vector u v. For
i; j 2 Z, sjv siu is the i applications of su followed by j
applications of sv . Therefore, for a point p 2 R2 ,
sjv siu p siujv p p iu jv. Two translations
su and sv are said to be independent if the two
vectors u and v are linearly independent.
A texture g is said to have a translation symmetry if there is a translation su such that g is
matched with su g in the non-empty overlapping
region (Stevens, 1991; Armstrong, 1988; Baglivo
and Graver, 1983; Martin, 1982). It is known that

1182

G. Oh, S. Lee / Pattern Recognition Letters 23 (2002) 11791190

the set of all such translations forms a group under

their composition (Armstrong, 1988). The group is
called the translation symmetry group of g and
denoted by Tg .
A texture g is regular if any translation in its
translation symmetry group Tg can be represented
by a repeated composition of two independent
translations, sa and sb (Washburn and Crowe,
1992; Armstrong, 1988; Baglivo and Graver, 1983;
Martin, 1982; Arnold, 1969). That is, g is regular if
two linearly independent vectors a and b exist,
such that Tg consists of translations sq , where
q ia jb; i; j 2 Z. Such a group Tg shall be
denoted by Tg ha; bi. Vectors a and b are called
the basis vectors of g. Note that a periodic texture
is not regular if it is periodic in only one direction.
Only textures that are periodic in two independent
directions can be termed regular textures.
For a point p 2 R2 , the lattice of g for p, denoted by Lg p, is the set of all points obtained by
applying the translations in Tg to p (Armstrong,
1988). That is,



Lg p qi;j jqi;j p ia jb; i; j 2 Z ;
where Tg ha; bi. When p is the origin o 0; 0,
we have



Lg o qi;j jqi;j ia jb; i; j 2 Z :
Note that Lg p is just a translation of Lg o by
sp . In this paper, we consider only the lattice
Lg o, which is called the lattice of g and denoted
by Lg . The origin o always belongs to Lg regardless of vectors a and b.
Depending on the shape of a parallelogram
determined by basis vectors a and b, we have
ve distinct lattice types: parallelogram, rectangle, rhombus, square, and hexagon (Washburn
and Crowe, 1992; Armstrong, 1988; Baglivo and
Graver, 1983):

Parallelogram. a and b do not belong to any of

the cases below.
Rectangle. a and b have dierent lengths, and
the angle between a and b is p=2.
Rhombus. a and b have the same length, and the
angle between a and b is neither p=3 nor p=2.
Square. a and b have the same length, and the
angle between a and b is p=2.
Hexagon. a and b have the same length, and the
angle between a and b is p=3.
Fig. 3 shows the congurations of the lattices.
Let g be a regular texture, such that Tg
ha; bi. A motif of g, denoted by Mg , is a parallelogrammic sub-region of g determined by its origin
m and basis vectors a and b. The origin m of Mg is
a point in g which determines a corner of Mg .
Then, Mg is specied by the parallelogrammic subregion of g whose vertices are m, m a, m b, and
m a b. See Fig. 4 for an illustration.
A motif Mg of a regular texture g contains the
minimal set of pixels in g such that we can reproduce g by repeating the pixel set. Although basis
vectors a and b have been determined, the extracted motifs may dier, depending on their origins. However, two motifs with dierent origins
have the same set of pixel information and only the
arrangements of pixels dier in the motifs. Any of
the motifs can be used to reproduce texture g by
tiling. Therefore, to determine a motif Mg , it is
sucient to identify basis vectors a and b. Now,
the problem of motif analysis of a regular texture g
is reduced to the problem of deriving two vectors a
and b, such that Tg ha; bi.

4. Basic approach
In this section, we rst present an approach to
analyze a motif of a given regular texture g. We

Fig. 3. Five lattice types: (a) parallelogram; (b) rectangle; (c) rhombus; (d) square; (e) hexagon.

G. Oh, S. Lee / Pattern Recognition Letters 23 (2002) 11791190

Fig. 4. Motif Mg of a regular texture g with the origin m.

then extend the approach to handle a noisy regular

texture f.
4.1. Motif analysis of a regular texture
Let g be a regular texture, whose lower left
corner is positioned at the origin o 0; 0. Let a
and b be basis vectors of g. That is, Tg ha; bi.
Then, the lattice Lg of g is given by Lg fqjq
ia jb i; j 2 Zg.
An important property of lattice Lg is that it
can be directly derived from texture g, even when
we do not know the basis vectors a and b. From
the denition of the translation symmetry group,
sq is a translation in Tg for each point q in Lg . In
other words, g is matched with sq g in the overlapping region for any point q in Lg . Hence, the
lattice Lg can be obtained by comparing g with its
translations and nding all points q, such that
g sq g.
After we determine the lattice Lg , the basis
vectors a and b can be derived from Lg . Let q be a
point in Lg and let Vg q be the set of vectors that
connect q to other points in Lg . Since the set
Vg q is the same regardless of the choice of q and
Lg always contains the origin o, we consider the
set Vg o, which is denoted by Vg . A pair of
vectors a and b in Vg can be basis vectors of g if a
and b satisfy the following properties (Armstrong,
1988; Galil and Park, 1992):
Vectors a and b are linearly independent.
The lengths of a and b are less than or equal to
those of other vectors in Vg ; that is, jjajj 6 jjvjj
and jjbjj 6 jjvjj for v 2 Vg fa; bg.
Refer to Fig. 3 for an illustration. Unfortunately,
such a pair of vectors in Vg is not unique. For

1183

example, if a and b satisfy the properties, so do

their negations, a and b.
To uniquely determine the basis vectors a and b
from Vg , we rst choose a as the vector in Vg ,
such that its angle from the positive x-axis is
minimum among the minimum length vectors in
Vg . Note that there are at least two minimum
length vectors in Vg , a and a. Then, b is assigned
to the vector in Vg , such that the angle between a
and b is minimum among the minimum length
vectors in Vg fa; ag. All the angles here are
measured in the counterclockwise direction. With
this selection scheme, a is always a vector in the
rst or second quadrant of the plane and the angle
between a and b is less than 180.
Now we can analyze a motif of a given regular
texture g in two steps: lattice determination and
basis vector determination. In the rst step, we
obtain the lattice Lg by searching the points q that
satisfy g sq g. The search is performed in the
region Rg , which contains the points p in R2 such
that the overlapped region of g and sp g is not
empty. When the size of g is m  n, we have
Rg fx; y j jxj < m and jyj < ng:
In the second step, we derive the vector set Vg
from the lattice Lg . Then, the basis vector a is the
minimum length vector in Vg , where we select the
vector whose angle from the x-axis is smaller
among the vectors with the same length. Similarly,
b is the minimum length vector in Vg fa; g,
where we consider the angle measured from a.
4.2. Motif analysis of a noisy regular texture
Let f be a noisy regular texture, which does not
satisfy the denition of a regular texture while still
preserving a visually regular structure. Texture f
can be regarded as a mixture of a regular texture g
and a random noise texture . That is, f g .
The problem of motif analysis of a given noisy
regular texture f can be dened as the detection of
two basis vectors a and b of the underlying regular
texture g. Then, a motif Mf of f can be extracted
from a parallelogrammic sub-region of f determined by the vectors a and b. By tiling the motif
Mf , we can generate a regular texture f 0 , which is
visually indistinguishable from f.

1184

G. Oh, S. Lee / Pattern Recognition Letters 23 (2002) 11791190

Let q be a point in the lattice Lg of g. Although

we have g sq g, it may not hold that f sq f
due to noise. To handle this problem, we use the
distance matching function df (Oh et al., 1999),
which computes the squared dierence between
f and sp f for a point p 2 Z2 . That is,
df p kf sp f k

regular texture f. Next, we present how to determine the threshold T, where the points p such that
df p 6 T are selected. The approximate lattice Lf
is obtained by choosing representative points in
the clusters of the selected points. Finally, we explain three types of errors that may exist in Lf .

2
2

kg sp g  sp k :
For a point q in Lg , df q is reduced to
2
k sq k . For texture f to preserve visually
regular structures, the magnitude of  should be
small over the entire domain. This implies that
function df would have a small value for each
point q in Lg . Hence, we can approximate lattice
Lg by using the function df .
Let Lf be an approximation of Lg , which may
contain errors. For example, Lf may contain extra
points that do not belong to Lg . Let Vf be the set
of vectors that connect the origin o to other points
in Lf . Because of the errors in Lf , we cannot
derive the basis vectors a and b of g by simply selecting minimum length vectors from Vf . For example, extra points in Lf may introduce vectors to
Vf whose lengths are smaller than those of a and b.
To resolve this problem, we use the Delaunay triangulation (de Berg et al., 1997) of Lf , which is the
dual of the Voronoi diagram of Lf . From the
edges of the Delaunay triangulation, we can obtain
a set of vectors that connect the nearby points in
Lf . The vectors a and b are then determined by
using the most frequent vectors in the set.
In summary, we analyze a motif of a given noisy
regular texture f in two steps: lattice approximation and basis vector approximation. In the rst
step, the approximate lattice Lf of f is obtained by
using the distance matching function df . In the
second step, we determine basis vectors a and b of f
by using the vector set derived from the Delaunay
triangulation of Lf . In Sections 5 and 6, we present the details of the rst and second steps, respectively.

5.1. Distance matching function

Let f be a texture of the size m  n. Let
p x; y be a point in Z2 . Then, the distance
matching function df is dened by
df p kf sp f k2
8
df1 x; y;
>
>
>
< d 2 x; y;
f

df1 x; y;
>
>
>
: d 2 x; y;
f

x P 0 and P 0;
x 6 0 and y P 0;
x 6 0 and y 6 0;
x P 0 and y 6 0;

where
df1 x; y

1
m xn y


m x 1
X n y 1
X
i0

df2 x; y

jf i; j f x i; y jj ;

j0

1
m xn y
m x 1
X n y 1
X

jf m 1 i; j
i0

j0
2

f m 1 i x; y jj :
As shown in Fig. 5, df p computes the average of
squared dierences of pixel values over the region
where f and sp f overlap.

5. Lattice approximation
In this section, we rst introduce the distance
matching function df (Oh et al., 1999) of a noisy

Fig. 5. Schematic diagrams of the distance matching function

df : (a) df1 x; y; (b) df2 x; y.

G. Oh, S. Lee / Pattern Recognition Letters 23 (2002) 11791190

The function value df p is well-dened at

points p x; y, jxj 0; 1; . . . ; m 1 and jyj
0; 1; . . . ; n 1, for which the overlapped region of
f and sp f is not empty. However, when jxj >
bm=2c or jyj > bn=2c, the overlapped region is less
than half of f, which makes the resulting function
values less reliable. Therefore, we only consider the
value of df at points p in Rf , where
n
jmk
Rf x; y j jxj 0; 1; . . . ;
and
j n2 ko
jyj 0; 1; . . . ;
:
2
By using the algorithm in (Oh et al., 1999), we can
compute the function values df p at all points
p 2 Rf in Omn log mn time.
Fig. 6 shows a noisy regular texture f scanned
from a book (Grafton, 1992) and the plot of the
corresponding df . In Fig. 6(b), the origin o lies at
the center of the image and the brightness at p is
proportional to the function value df p. Note that
function values df p have the same regular structure as texture f.
5.2. Approximate lattice determination
Let f be an m  n noisy regular texture such that
f g , for a regular texture g and a noise .
Since f contains , the function value df p may not
be zero for a point p in the lattice Lg of g. To
approximate Lg , we collect all points q in Rf such
that df q is below a given threshold T. That is,



L0f qjdf q 6 T ; q 2 Rf :
If the threshold T is large, every point in Lg
could be included in L0f , but we may have many
extra points in L0f which do not belong to Lg .

Fig. 6. A texture f and its distance matching function df : (a)

texture f; (b) the plot of df .

1185

When we select a smaller T, the number of extra

points in L0f decreases, but L0f can miss points in
Lg .
To obtain a reasonable threshold T, we consider the expected value ld of df p for an arbitrary
point p x; y in Rf . We may assume that the
pixel values of f are integers between 0 and Q. Let
h be the histogram of f, where hc gives the
number of pixels in f whose values are c. Consider
a pixel q i; j in f and let c represent its pixel
value. Pixel q is compared to the pixel q0
i x; j y in computing df p. The probability
that pixel value f q0 is d can be approximated by
hd=mn. Hence, the expected squared
PQdierence
between pixel values of q and q0 is
d0 hd=
mnc d2 . Then, the expected value ld of df p
can be obtained by weighted averaging these expected squared dierences over all pixel values c.
That is,
ld

Q
Q
X
1 X
hd
2
c d
hc
mn c0
mn
d0
Q
Q
1 XX
2
hchdc d :
m2 n2 c0 d0

For a point q in Lg , function value df q would be

much less than ld , which is the expected value of
df p for an arbitrary point p in Rf . We introduce
the noisiness parameter a between 0 and 1, and use
ald as the threshold T to obtain L0f . When f appears noisy, we use a larger a, so that L0f contains
as many points in Lg as possible. A smaller a
would be sucient for a less noisy f. In our experiments, a between 0:2 and 0:5 worked well for
most noisy regular textures.
When a relatively large threshold T is used, the
point set L0f may contain many extra points that
do not reside in Lg . For example, Fig. 7(a) shows
the plot of L0f obtained by applying the threshold
T with a 0:5 to the function values df p in Fig.
6(b). As we can see in Fig. 7(a), the points in L0f
tend to form clusters around the points in Lg . To
rene L0f , we decompose L0f into point clusters
and select a representative point from each cluster.
Let Si , i 0; 1; . . . ; l 1, represent the point
clusters in L0f , which satisfy the following conditions:

1186

G. Oh, S. Lee / Pattern Recognition Letters 23 (2002) 11791190

Fig. 7. Example of lattice approximation: (a) point set L0f ; (b)

approximate lattice Lf .

Sl 1
L0f i0 Si and Si \ Sj ; for all i 6 j.
For all p; q 2 Si , there exists an 8-connected
path in Si from p to q.
The rst condition implies that the set of clusters
Si is a partition of L0f . With the second condition,
each cluster Si is an 8-connected component of
L0f . In each cluster Si , we choose a point qi whose
function value df qi is the minimum among those
of the points in Si . We can consider the set of such
points qi as an approximation of the lattice Lg .
The point set is denoted by Lf and called the
approximate lattice of f:


Lf qi jdf qi 6 df q for all q 2 Si ;

i 0; 1; . . . ; l 1 :
For a given noisy regular texture f of size m  n,
the approximate lattice Lf can be obtained in
Omn log mn time, because the evaluation of the
distance matching function df dominates the
computation time. Fig. 7(b) shows the plot of Lf
derived from L0f shown in Fig. 7(a). From Fig.
7(b), we observe that the clustering process eectively eliminates extra points in L0f .
5.3. Possible errors in approximate lattice Lf
Let f be a noisy regular texture and g the underlying regular texture. The approximate lattice
Lf of f may be dierent from the lattice Lg of g.
This is due to three types of errors: perturbation
errors, missing points, and extra points. With a
perturbation error, the position of a point q in Lg
changes slightly in Lf . This error occurs when
point q does not have the minimum value of
function df among the points in the cluster to
which it belongs. A point q in Lg can be missed in

Lf if there is much noise in the overlapping region

of f and sq f , so that df q is greater than the
threshold T. An extra point q may exist in Lf ,
which is not contained in Lg , if the noise in the
overlapping region of f and sq f accidentally reduces df q to below T.
Among these three types of errors, we can
almost entirely eliminate the missing points by
taking a relatively large threshold T. In the experiments, threshold T ald with a 0:5 usually
resulted in approximate lattices Lf with no missing points. In the next section, we present a robust
method to determine the basis vectors a and b of g
from Lf , which properly handles perturbation
errors and extra points.

6. Basis vector approximation

In this section, we rst investigate the property
of the Delaunay triangulation of a lattice Lg of
a regular texture g. Then, the property is applied
to the approximate lattice Lf of a noisy regular texture f to determine the basis vectors a and
b of f.
6.1. Delaunay triangulation of a lattice
Let g be a regular texture. As mentioned in
Section 3, the lattice Lg of g belongs to one of the
ve lattice types; parallelogram, rectangle, rhombus, square, and hexagon. Fig. 8 shows the Delaunay triangulations of Lg for the ve lattice types.
Note that in the Delaunay triangulation of a point
set, there are edges only between nearby points (de
Berg et al., 1997).
Let Dg be the Delaunay triangulation of Lg .
Regardless of the lattice type of Lg , there are six
vectors in Dg which connect the origin o to its
neighboring points in Lg (see Fig. 8). Recall that o
always belong to Lg . These vectors can be represented by three linearly independent vectors, c1 , c2 ,
and c3 , and their negations. Without loss of generality, we may assume that 0 6 \c1 < \c2 <
\c3 < 180, where \ci denotes the angle of ci
measured from the positive x-axis. As we can see
in Fig. 8, the vectors ci satisfy that c3 c2 c1 for
all lattice types.

G. Oh, S. Lee / Pattern Recognition Letters 23 (2002) 11791190

1187

Fig. 8. Delaunay triangulations for ve lattice types: (a) parallelogram; (b) rectangle; (c) rhombus; (d) square; (e) hexagon.

Let a and b represent two vectors among c1 ; c2 ,

and c3 , such that jjajj 6 jjci jj for all ci and jjbjj 6
jjci jj for ci 6 a. Then, a and b satisfy the property
of the basis vectors of g specied in Section 4.1.
That is, a and b are linearly independent, and their
lengths are less than or equal to those of other
vectors in Vg , where Vg is the set of vectors from
o to other points in Lg . This implies that we are
not required to consider all vectors in Vg to determine the basis vectors of g. Instead, it is sucient to consider only six vectors in C
fc1 ; c2 ; c3 g. Based on the selection scheme
outlined in Section 4.1, we choose the basis vectors
a and b of g as the minimum length vectors in C
and C fa; ag whose angles from the x-axis and
a are the smallest, respectively.
Although the vectors ci are obtained from the
edges in Dg which connect o to its neighbor points,
they can also be used to represent the other edges
in Dg . Since Dg has a regular structure, for any
interior point q in Dg , an edge from q to its nearby
point corresponds to one of the six vectors in C
(see Fig. 8). That is, the set of all interior edges in
Dg can be reduced to the vector set C. We use this
property to determine the three vectors, c1 ; c2 , and
c3 , from the approximate lattice Lf of a noisy
regular texture f.
6.2. Basis vector determination
Let f be a noisy regular texture, such that
f g , for a regular texture g and a noise . Let
Df be the Delaunay triangulation of the approximate lattice Lf . Due to the errors in Lf , the set of
interior edges in Df may not be represented by
three vectors, c1 , c2 , and c3 . For example, an extra

point in Lf introduces an edge to Df which has no

relation with ci . Also, with the perturbation errors
in Lf , slightly dierent vectors are extracted from
the edges in Df for which would result in the same
vector ci if there were no perturbation errors.
Hence, we determine c1 ; c2 , and c3 by nding the
three largest clusters in the vectors obtained from
the edges in Df and choosing representative vectors in the clusters.
Let e be an edge in Df that connects two
points q and r. Without loss of generality, we assume that qx < rx if qy ry , and qy < ry otherwise,
where q qx ; qy and r rx ; ry . Then, edge e is
considered as the vector from q to r and represented by the point rq in the plane. With
this conversion, each edge in Df can be mapped
to a point in the rst or second quadrant of the
plane.
Let Ef be the set of such points derived from the
interior edges in Df . We decompose Ef into a set
of point clusters Si by the same method we performed for L0f in Section 5.2. Next, we choose
the three largest clusters from Si in terms of the
number of points in each cluster. By averaging the
points in each of these three clusters, we obtain
three vectors c01 , c02 , and c03 , where 0 6 \c01 <
\c02 < \c03 < 180.
Since lattice Lf may contain perturbation
errors, the vectors c0i do not necessarily satisfy c03
c02 c01 . We determine the vectors c1 ; c2 , and c3
from c0i by minimizing the least-square error
1
errc1 ; c2 kc1 c01 k2 kc2 c02 k2
2
2

kc2 c1 c03 k :

1188

G. Oh, S. Lee / Pattern Recognition Letters 23 (2002) 11791190

Vectors c1 and c2 that minimize errc1 ; c2 can

be obtained by solving the linear system of equations

then give several examples of motif analysis of

noisy regular textures.
7.1. Motif extraction

oerr
2c1 c2 c01 c03 0;
oc1
oerr
c1 2c2 c02 c03 0:
oc2
From the equations, we nally obtain
2c01 c02 c03
;
3
c3 c2 c1 :

c1

c2

c01 2c02 c03

;
3

As mentioned in Section 5.3, lattice Lf may contain extra points and perturbation errors if we
employed a large threshold T to obtain the
point set L0f . When we determine the vectors ci
from Lf with the above approach, the eects of
the extra points are properly removed by selecting the three largest clusters from the point set
Ef . The perturbation errors are handled by averaging the points in the selected clusters. The leastsquare tting of c1 and c2 further enhances the
robustness of the method against perturbation
errors.
The Delaunay triangulation of a point set with
p points can be constructed in Op log p time (de
Berg et al., 1997). Let m  n be the size of a given
noisy regular texture f. Then, lattice Lf contains
at most mn points and the Delaunay triangulation
Df of Lf can be obtained in Omn log mn time.
This implies that we can obtain the three vectors,
c1 ; c2 , and c3 , from Lf in Omn log mn time because the construction of Df dominates the computation time. The basis vectors a and b of f can
be determined from C fc1 ; c2 ; c3 g, as mentioned in Section 6.1.

7. Motif analysis examples

In this section, we rst explain in detail how to
extract a motif Mf from a noisy regular texture f
when we know the basis vectors a and b of f. We

Let f be a noisy regular texture and a and b be

the basis vectors of f determined by the techniques
proposed in Sections 5 and 6. To obtain a motif
Mf of f by using a and b, we must select the origin m of Mf among the points in f. Although
other choices are possible, we select m mx ; my
such that mx minf0; ax ; bx ; ax bx g and my
minf0; by g, where a ax ; ay and b bx ; by .
Recall that ax , bx , and by may have negative
values. With this choice of m, the motif Mf , which
is the parallelogrammic sub-region determined
by m, a, b, and a b, is guaranteed to be contained
in f.
After origin m has been selected, motif Mf can
be generated by collecting pixels p of f in the
parallelogrammic sub-region. That is,
Mf p f m p for p k1 a k2 b;
where 0 6 k1 ; k2 < 1:
By tiling Mf in the directions of a and b, we obtain
a regular texture f 0 of an arbitrary size, which is
visually indistinguishable from f.
7.2. Examples
Fig. 9(b) shows the analyzed motif of the texture in Fig. 9(a), which is the same as in Fig. 6(a).
By tiling the motif, we obtain the texture in Fig.
9(c), which is larger than the original. The texture
in Fig. 9(d) was scanned from a book (Hornung,
1976). Its motif and a texture generated by tiling
the motif are shown in Figs. 9(e) and (f), respectively. The texture in Fig. 9(g) was obtained from
a web site (Schoeld, 2001) and is commonly referred as the Eschers design. Figs. 9(h) and (i)
show its motif and a texture generated by tiling the
motif, respectively. The texture in Fig. 9(j) was
obtained by scanning a texture in a book (Stegenga, 1992). Figs. 9(k) and (l) show a motif and a
texture generated by tiling the motif, respectively.
For a given m  n noisy regular texture, lattice
and basis vector approximations can be performed

G. Oh, S. Lee / Pattern Recognition Letters 23 (2002) 11791190

1189

Fig. 9. Motif analysis examples of noisy regular textures.

in Omn log mn time, as described in Sections 5

and 6, respectively. Hence, we can obtain the basis
vectors a and b from f in Omn log mn time. The
sizes of the example textures in Fig. 9, from top to
bottom, are 273  179, 210  235, 209  284, and
208  187, respectively. For the textures, the computation times for the motifs shown in Fig. 9 are

9, 5, 12, and 6 seconds on an SGI Indigo2 with a

R10000 processor, respectively.
8. Conclusion
A regular texture contains spatial redundancy
due to its translation symmetry. By contrast, a

1190

G. Oh, S. Lee / Pattern Recognition Letters 23 (2002) 11791190

motif of the texture avoids such redundancy and

only includes a minimal pattern that can reproduce the texture by tiling. In this respect, a motif
gives a more compact representation of a regular
texture without losing structural information.
In this paper, we proposed an algorithm to
obtain a motif of a noisy regular texture in Omn
log mn time, where m  n is the size of the texture.
This algorithm is applicable to texture analysis,
texture recognition, and texture compression.
Acknowledgements
This work was supported in part by the Korea
Ministry of Education through the Brain Korea 21
program, KIPA Game Animation Research Center, and the KOSEF Virtual Reality Research
Center.
References
Armstrong, M.A., 1988. Groups and Symmetry. Springer, New
York.
Arnold, F., 1969. The Geometry of Regular Repeating Pattern.
The Textile Institute.
Baglivo, J.A., Graver, J.E., 1983. Incidence and Symmetry in
Design and Architecture. Cambridge University Press,
Cambridge, MA.
Conners, R.W., Harlow, C.A., 1980. Toward a structural
analyzer based on statistical methods. Comput. Graphics,
Image Process. 12, 224256.

de Berg, M., van Kreveld, M., Overmars, M., Schwarzkopf, O.,

1997. Computational Geometry: Algorithms and Applications. Springer, Berlin.
Galil, Z., Park, K., 1992. Truly alphabet-independent twodimensional pattern matching. In: Symposium on Foundations of Computer Science FOCS, pp. 247256.
Galil, Z., Park, K., 1996. Alphabet-independent two-dimensional witness computation. SIAM J. Comput. 25 (5), 907
935.
Grafton, C.B., 1992. Decorative Tile Designs. Dover Publications, New York.
Hornung, C.P., 1976. All over Patterns for Designers and
Craftsmen. Dover Publications, New York.
Martin, G.E., 1982. Transformation Geometry. Springer,
Berlin.
Oh, G., Lee, S., Shin, S.Y., 1999. Fast determination of textural
periodicity using distance matching function. Pattern Recognition Lett. 20 (2), 191197.
Parkkinen, J., Selkainaho, K., 1990. Detecting texture periodicity from the co-occurrence matrix. Pattern Recognition
Lett. 11, 4350.
Porter, A.W., 1975. Principles of Design Pattern. Davis
Publications Inc.
Schoeld, P., 2001. Available from http://www.worldofescher.
com.
Stegenga, W., 1992. Pictorial Archive of Geometric Designs.
Dover Publications, New York.
Stevens, P.S., 1991. Handbook of Regular Patterns An
Introduction to Symmetry in Two Dimensions. MIT Press,
Cambridge, MA.
Washburn, D.K., Crowe, D.W., 1992. Symmetries of Culture:
Theory and Practice of Plane Pattern Analysis. University
of Washington Press.
Zucker, S.W., Terzopoulos, D., 1980. Finding structure in cooccurrence matrices for texture analysis. Comput. Graphics,
Image Process. 12, 286308.