Journal of Theoretical and Applied Information Technology

28th February 2014. Vol. 60 No.3

© 2005 - 2014 JATIT & LLS. All rights reserved.

ISSN: 1992-8645 www.jatit.org E-ISSN: 1817-3195

AUTOMATIC MOTIF DETECTION FOR ISLAMIC

GEOMETRICAL PATTERNS
1
ABDELBAR NASRI, 2RACHID BENSLIMANE
1,2
Laboratoire de Transmission et de Traitement de l’Information
Université Sidi Mohamed Ben Abdellah, USMBA, Ecole supérieure de Technologie
Route d’Imouzzer BP. 2427, Fès, Maroc
E-mail: 1 [email protected], 2 [email protected]

ABSTRACT

In this paper, we present a new method to detect the basic unit cell of a periodic Islamic Geometrical
Pattern (IGP). This method is based on the autocorrelation function (ACF), a function known to be
appropriate to analyze and extract a repetitive motif from a regular texture. The motif can be successfully
extracted when the peaks detected in the autocorrelation function of an image are pertinent. To optimize the
peaks detection, we propose a new method based on the stability of the motif surface which is defined by
two short displacement vectors. Compared to classical extraction methods of periodicity, the proposed
method hereby is tolerant to geometric distortion, noise and changes in intensity. Tests on 166 images with
different visual quality demonstrate the capability of the proposed method to extract the periodic motif
automatically without the need of human intervention.
Keywords: Autocorrelation, Wallpaper Groups, Islamic Art, Pattern Extraction, Displacement Vector.

1. INTRODUCTION

Texture periodicity analysis has become popular

lately and has been used for texture tracking [14],
synthesis [22, 11, 31], and retrieval [8, 9].
It is proven that every 2D periodic patterned
texture can be classified in one of five groups based
on the type of network or lattice (unit cell or motif)
as illustrated in figure (1).
Previous works proposed for motif extraction can
be grouped broadly into two categories: the local
feature-based approach [6, 13, 14, 15, 16, 17, 18,
19, 25], and the global structure-based approach [1,
3, 7, 8, 9, 20, 21, 22, 23].
The local feature-based approach starts by Image used in (1, 2, 3) are captured from [29]
identifying a number of texel candidates. Matching
based on visual similarity between these potential
(1): Parallelogram, t1 ≠ t2 , ϕ ≠ 90 ,
o

texels and their neighbors is then performed.

(2): Rectangular, t1 ≠ t2 , ϕ = 90
Successful matching leads to the connection of o

texels with a lattice structure. The approach is

repeated until no more new texels are found. (3): Centered rectangular t1 ≠ t2 , ϕ = 90
o

The global structure-based approach [1, 3, 4, 7, 8,

(4): Hexagonal, t1 = t2 , ϕ = 6 0
o
9, 20, 21, 22, 23.] tries to extract periodicity using
methods that emphasize the idea of periodic
(5): square, t1 = t 2 , ϕ = 9 0
o
patterns as global processes.
T.Matsuyama [4] proposes an approach based on Figure 01: The Five Fundamental Two-Dimensional
Fourier transformation. The author of [5, 23.] Bravais Lattices.

594
 Journal of Theoretical and Applied Information Technology
28th February 2014. Vol. 60 No.3
© 2005 - 2014 JATIT & LLS. All rights reserved.

ISSN: 1992-8645 www.jatit.org E-ISSN: 1817-3195

presents methods that use periodicity measures

defined by co-occurrence matrices computed
globally over the whole pattern. A: ACF

A more traditional image processing approach to

detect repetition uses autocorrelation function ACF Source image (a)
[1, 3, 7, 20, 21.]. The main problem with using this
approach is the difficulty of finding the pertinent C: Displacement B: Pertinent
vectors peaks
peaks in the ACF.
Chetverikov [20] proposes a regularity measure (c)
(b)
based on finding the maximum overall direction on
the ACF. Leu [21] uses the several highest peaks in Motifs
(d)
the ACF, computed on the gradient field of the extraction
image to capture translation vectors.
H.C. Lin et al [1] present a threshold-free Figure 2: Process To Extract Periodic Motif Based On
approach based on a filtered ACF and obtained by ACF
using an optimal Gaussian smoothing. Yanxi Liu et
al [7] propose an efficient method based on the A. Autocorrelation function
region of dominance. All these methods select
arbitrarily the number of peaks in the ACF. To
overcome this drawback, Jingrui He et al [3] Lin et al [1] prove that if a grey-scale
propose an optimization scheme to automatically image I(x, y) presents a periodic motif, the shape
extract the pertinent peaks and the correct number and arrangement of this motif can be found using
of peaks. Despite its good performance, this method the locations of the peaks in its autocorrelation
is sensible to geometrical distortion, noise and function ac(x,y), computed as follows:
quality of the processed images.
A B
In this paper we propose a new automatic method ac ( x, y ) = ∑∑ I * (i, j ) I ( x + i, y + j ) (1)
based on ACF. This method is rapid and no i =1 j =1
sensitive to geometric distortions, noises and
intensity changes.
Where " *" is the complex conjugate and
The rest of this paper is organized as follows: A and B define the dimension of the original image.
Section 2 describes an approach to extract motifs According to the correlation theorem Gonzalez and
based on the autocorrelation function. Section 3 Woods [26], Eq. (1) can be expressed as:
presents the proposed method. Experimental results
ac ( x, y ) = F −1  F * ( I ( x, y)) F ( I ( x, y))  (2)
of these latter are presented in Section 4. Finally,
conclusions are drawn in Section 5.
2. PERIODICITY EXTRACTION BASED ON
−1
ACF: Where F and F denote respectively
the Fourier transform and inverse Fourier
The process to extract periodicity with transform.
ACF consists of three major steps: (A) Calculation
of the Autocorrelation function, (B) Finding of B. Finding pertinent peaks
pertinent peaks, and (C) Displacement of vectors
and extraction of the motif. Figure 2 gives an
Pertinent peaks detection follows generally
illustration of these steps.
two steps:

• A coarse peaks detection step: ac(x,y) is

considered as a peak if it is the maximum of all
the ac(x0, y0), where (x0,y0) is a neighbor of
(x,y) in a window of fixed size VxV with the
center (x,y). The number of peaks detected in
this step is distorted by false peakes.

595
 Journal of Theoretical and Applied Information Technology
28th February 2014. Vol. 60 No.3
© 2005 - 2014 JATIT & LLS. All rights reserved.

ISSN: 1992-8645 www.jatit.org E-ISSN: 1817-3195

• A fine peak detection step allowing to find utilized to find the two displacement vectors [1].
only pertinent peaks. There exist manual and These two non-collinear vectors are used as two
automatic methods. translation vectors that generate the lattice, from
which the repetitive motif is extracted (cf. Fig 3).
• Manual peak number detection
Lin et al [1] present a threshold-free approach
based on finding an optimal Gaussian smoothing of
the autocorrelation function. To obtain a smoothed
autocorrelation function, they convolve the 2D ACF Unit cell
with a 2D Gaussian filter (Eq (3)).

1  − x2 − y 2 
g ( x, y , σ ) = exp  2σ 2  Eq (3)
2πσ 2  

As it is proposed by Yianx et al [7] “the

absolute height of a peak is not as important as the
size of its region of dominance (RD), defined as the
largest circle centered on the candidate peak such
that no higher peaks are contained in the circle. A
Figure 3: Unit Cell Or Lattice Surrounded By 2
peak with a low height, but located far from any uuur
Calculated Displacement Vectors PP1 And PP2
uuur
larger neighbors, is much more perceptually
important than a high peak that is close to an even
higher one. The first 32 most dominant peaks found 3. PROPOSED METHOD
using this method are well distributed over the
whole image, with very few spurious peaks”. We propose an iterative method to
automatically extract the repetitive motif that
This method is good at removing false identifies Islamic Geometry patterns: The optimal
peaks, especially in an IGP that contain a large number of peaks and the corresponding two
number of repetitive patterns (cf. Fig A.2). displacement vectors are determined from the
However, with this method, the number of pertinent autocorrelation function.
peaks has to be selected manually. Moreover, it
does not detect pertinent peaks in the IGP in the By varying the number N of peaks, which
case of a repetitive pattern taking a large proportion are the N highest peaks in the ACF, we calculate
of the processed image. (cf. Fig A.2) illustrates an the most frequent surface ‘S’ of the periodic motif,
example of our reimplementation of this method. defined by two displacement vectors.

• Automatic peak number detection This iterative procedure stops when the
Jingrui He et al [3] propose an surface ‘S’ of the periodic motif remains unchanged
optimization-based scheme for automatic peak between two successive iterations n and n+1,
number detection. They first formulate it as an S(n)=S(n+1). This iterative process results in the
optimization problem by a straightforward yet determination of the optimal number of peaks and
effective criterion function, and then resort to two linearly independent translation vectors that
Simulated Annealing to optimize it. Despite its good generate the lattice, from which the periodic motif
performance, this method needs an adaptive is extracted.
parameterization based on some prior knowledge of
the database [3]. Figure A.3 illustrates an example The main procedure is to determine for a
of automatic peak detection using our chosen number of peaks N the most frequent two
reimplementation of this method. displacement vectors. To do so, we proceed as
follows:
C. Displacement vectors
• Compute the autocorrelation function as
Once peak number is determined, the mentioned in section 2-A.
Generalized Hough Transform (GHT) [1] can be • Sort peaks in descending order of magnitude.

596
 Journal of Theoretical and Applied Information Technology
28th February 2014. Vol. 60 No.3
© 2005 - 2014 JATIT & LLS. All rights reserved.

ISSN: 1992-8645 www.jatit.org E-ISSN: 1817-3195

• Repeat while S(n)==S(n+1). The performance of the proposed method

• Select the N highest peaks is evaluated by comparing its correct rate and time
• For each highest peak P(x,y) do: complexity with those proposed by Jingrui He et al
o Find the two nearest neighbor peaks [3] and by Lin [1]. 166 images were used to test
P1 ( x1 , y1 ) and P2 ( x2 , y2 ) with the constraint this performance. Some representative unit cells are
shown in Figure (A.8). The tested images have
that the angle between the resulting vectors
uuur uuur different photometric and geometric quality. They
PP1 and PP2 must be between 60° and also have different sizes varying from 128x128
90° [28] (cf. Fig 3). pixels to 1700x1300 pixels.
o Compute the surface of the unit cell (motif)
surrounded by the two calculated Figure (A.5, A.7) shows some examples,
for which our method leads to good detection in
uuur uuur
displacement vectors PP1 and PP2 (cf.
comparison with Jingrui’s method which fails to
Fig 4). detect the right motifs (cf. Fig A.4, A.6).
o Compute the most frequent displacement
vectors using their corresponding surface Table 1 shows that the proposed method
motif S(n). presents the best rate of the correct detection of the
periodic motifs compared to Jingrui He et al
method.

Table 2 shows for the first 5 images of

figure 8, that our method is more rapid that Jingrui
et al method. In fact, our method calculates surface
N times in each iteration ο (n) , and method [3]
calculates SA, NxN times in each iteration of
algorithm ο (n 2 ) , the difference is more important
when the number of peaks N increases.

Table 1: Comparison Between [3] And Our Method As A

Figure 4: This Figure Shows An Example For N=12. Correct Rate
This Figure Plot The Surface Of Unit Cell Calculated In Method Correct
Correct False
Each Peak And The Most Occurred Surface. detection
detection detection
Rate
Jingrui He et al [3] 143 23 86.14%
Our processes to
155 11 93.37%
extract periodicity

Table 2: Comparison Between [3] And Our Method As A

Time Complexity
Source images Jingrui He Our processes to
et al [3] extract periodicity
Image 1
1.5130 s 1.3246 s
587x384
Image 2
4.2081 s 0.7243 s
430x323
Image 3
4.0677 s 1.9096 s
Figure 5: This Figure Shows That The Optimal Peak 600x480
Number Is N=16, Because The Surface Remains Image 4
5.2447 s 1.7013 s
Unchanged From N=16 To N=17. 645x419
Image 5
2.5080 s 2.0945 s
610x467

5. CONCLUSION

4. EXPERIMENTAL RESULTS

597
 Journal of Theoretical and Applied Information Technology
28th February 2014. Vol. 60 No.3
© 2005 - 2014 JATIT & LLS. All rights reserved.

ISSN: 1992-8645 www.jatit.org E-ISSN: 1817-3195

In this paper, a motif detection method for and Machine Intelligence, VOL. 26, NO. 3,
periodic Islamic patterns based on the MARCH 2004
autocorrelation function was proposed. The test [8] Junwei Han, Stephen J. McKenna, Classifying
results show the robustness of the proposed method And Comparing Regular Textures For Retrieval
to extract periodicity from images with different Using Texel Geometry, Proceeding Of Visapp,
photometric and geometric quality, and with Portugal, 5-8 February, 2009.
different size. [9] Junwei Han, Stephen J. McKenna, and Ruixuan
Wang, Regular Texture Analysis as Statistical
The following conclusions can be drawn Model Selection, Springer-Verlag Berlin
from the proposed method: Heidelberg 2008,ECCV 2008, Part IV, LNCS
5305, pp. 242–255, 2008.
• The method is simple in implementation and [10] Leung, T., Malik, J.: Recognizing surfaces
flexible in synthesizing the regular texture. using three-dimensional textons. In: IEEE
• There is no need for pre-processing or post International Conference on Computer Vision,
processing such us filtering and thresholding Corfu, Greece, pp. 1010–1017 (1999)
of input image. [11] Liu, Y., Tsing, Y., Lin, W.: The promise and
• The proposed method can be used to perils of near-regular texture. Inter-national
automatically detect periodicity in textures Journal of Computer Vision 62, 145–159 (2005)
belonging to 17 wallpaper groups. [12] Malik, J., Belongie, S., Shi, J., Leung, T.:
Textons, contours and regions: cue integration
REFRENCES: in image segmentation. In: IEEE International
Conference of Computer Vision, Corfu, Greece,
[1] H.-C. Lin, L.-L. Wang, and S.-N. Yang, pp. 918–925 (1999).
“Extracting Periodicity of a Regular Texture [13] Hays, J., Leordeanu, M., Efros, A., Liu, Y.:
Based on Autocorrelation Functions”, Discovering texture regularity as a higher-order
ELSEVER, Pattern Recognition Letters,vol. 18, correspondence problem. In: European
pp. 433-443, 1997. Conference on Computer Vi-sion, Graz,
[2] V. Asha, P. Nagabhushan, N.U. Bhajantri, Austria, pp. 533–535 (2006)
Automatic extraction of texture-periodicity [14] Lin, W., Liu, Y.: A lattice-based MRF model
using superposition of distance matching for dynamic near-regular texture tracking.IEEE
functions and their forward differences, Transactions on Pattern Analysis and Machine
ELSEVIER, Pattern Recognition Letters 33 Intelligence 29, 777–792 (2007).
(2012) 629–640. [15] Leung, T., Malik, J.: Detecting, localizing and
[3] Jingrui He and al, Automatic Peak Number grouping repeated scene elements from an
Detection in image Symmetry analysis, PCM image. In: European Conference on Computer
2004, LNCS 3333, pp. 111-118, 2004. Springer- Vision, Cambridge, UK, pp. 546–555 (1996).
Verlag Berlin Heidelberg 2004. [16] Tuytelaars, T., Turina, A., Gool, L.:
[4] T.Matsuyama, S.Miura, and M. Nagao. A Noncombinational detection of regular repeti-
structural analysis of natural textures by fourier tions under perspective skew.IEEE Transactions
transformation. CVGIP, 24(3):347-362, on Pattern Analysis and Machine Intelligence
December 1983. 25, 418–432 (2003).
[5] R.W. Conners and C.A. Harlow, Toward a [17] Schaffalitzky, F., Zisserman, A.: Geometric
structural textural analyzer based on statistical grouping of repeated elements within images.
methods. CGIP, 12(3):224-256, March 1980. In: Shape, Contour and Grouping in Computer
[6] Manuel Agusti-Melchor et al, lattice extraction Vision. Lecture Notes in Computer Science, pp.
based on symmetry analysis, VISAPP 2008- 165–181.Springer, Heidelberg (1999).
international conference on computer vision [18] Forsyth, D.A.: Shape from texture without
theory and applications 2008. boundries. In: European Conference in
[7] Yanxi Liu, Robert T. Collins and Yanghai Tsin, Computer Vision, Copenhagen, Denmark, pp.
A Computational Model for Periodic Pattern 225–239 (2002).
Perception Based on Frieze and Wallpaper [19] Lobay, A., Forsyth, D.A.: Recovering shape and
Groups, IEEE Transactions on Pattern Analysis irradiance maps from rich dense texton fields.

598
 Journal of Theoretical and Applied Information Technology
28th February 2014. Vol. 60 No.3
© 2005 - 2014 JATIT & LLS. All rights reserved.

ISSN: 1992-8645 www.jatit.org E-ISSN: 1817-3195

In: Computer Vision and Pattern Recognition,

Washington, USA, pp. 400–406 (2004).
[20] Chetverikov, D.: Pattern regularity as a visual
key. Image and Vision Comput-ing 18, 975–985
(2000)
[21] Leu, J.: On indexing the periodicity of image
textures. Image and Vision Comput-ing 19,
987–1000 (2001)
[22] Charalampidis, D.: Texture synthesis: Textons
revisited. IEEE Transactions on Image
Processing 15, 777–787 (2006)
[23] Starovoitov, V., Jeong, S.Y., Park, R.: Texture
periodicity detection: features, properties, and
comparisons. IEEE Transactions on Systems,
Man, and Cybernetics-A 28, 839–849 (1998)
[24] Gruenbaum, B. et al, Tilings and patterns. W
.H. free-man and company, New Yourk (1987).
[25] Yan Gui and al, Texel based regular and Near-
regular Texture Characterization, International
conference of multimedia and signal processing
2011.
[26] Gonzalez, R.C. and R.E. Woods (1993). Digital
Image Processing. Addison-Wesley, MA,
3rded.
[27] Lin, H.-C., L.-L. Wang and S.-N. Yang (1996).
Automatic deter- mination of the spread
parameter in Gaussian smoothing. Pattern
Recognition Letters 17 (12), 1247-1252.
[28] D. Schattschneider.The plane symmetry
groups:their recognition and notation.American
Mathematical Monthly, 85:439, 1978.
[29] J H Conway, H Burgiel and C Goodman-
Strauss. The symmetry of things. A K Peters
Ltd. 2008. ISBN 978 1 56881 220 5.
[30] G Loy and J. Eklundh, “Detecting symmetry
and symmetric constellations of features."
inProc. European Conf. Computer Vision, 2006,
pp. 508–521.
[31] T.Gangopadhyay XLRI C.H.Area(E),
Jamshedpur, Further Tiling Patterns Involving
Islamic Rosettes with an Odd Number of
Vertices, International Journal of Computer
Applications (0975 – 8887) Volume 71– No.6,
May 2013

599
 Journal of Theoretical and Applied Information Technology
28th February 2014. Vol. 60 No.3
© 2005 - 2014 JATIT & LLS. All rights reserved.

ISSN: 1992-8645 www.jatit.org E-ISSN: 1817-3195

Source image without RD with RD

Figure A. 2: Dominant Peaks Found Using Region Of Dominance RD.

(a) (b) (c)

Figure A. 3: Example Of Automatic Peak Number Detection (B) Via Optimizing The Criterion Function (C), In
This Example Nopt=19.

C(N)
Peak number

Figure A. 4: Example Of Automatic Peak Number Detection Via Optimizing The Criterion Function C(N) [3].
Surface
Peak number

Figure A. 5: Example Of Images For Which The Proposed Method Extracts The Correct Unit Cell, While The
Method Of Jingrui Et Al. Falls To Extract Them.

600
 Journal of Theoretical and Applied Information Technology
28th February 2014. Vol. 60 No.3
© 2005 - 2014 JATIT & LLS. All rights reserved.

ISSN: 1992-8645 www.jatit.org E-ISSN: 1817-3195

C(N)

Peak number
Figure A. 6: Example Of Falls Unit Cell Detection Using Method Of [3].
Surface

Peak number

Figure A. 7: Correct Unit Cell Using Our Method

Figure A. 8: Results Of Automatically Detected Lattices Using Our Method.

601