Colour-Texture Image Segmentation using Hypercomplex Gabor Analysis

Signal & Image Processing : An International Journal(SIPIJ) Vol.1, No.2, December 2010
DOI : 10.5121/sipij.2010.1207 75
Colour-Texture Image Segmentation using
Hypercomplex Gabor Analysis
B.D.Venkatramana Reddy1
and Dr.T.Jayachandra Prasad2
1
Department of Electronics and Communication Engineering, Madanapalle Institute of
Technology&Science, Madanapalle-517325, Andhra Pradesh, India
balam.diguva@gmail.com
2
Department of Electronics and Communication Engineering, RGM College of
Engineering&Technology, Nandyal-518501, Andhra Pradesh, India
jp.talari@gmail.com
Abstract
Texture analysis such as segmentation and classification plays a vital role in computer vision and pattern
recognition and is widely applied to many areas such as industrial automation, bio-medical image
processing and remote sensing. In this paper, we first extend the well-known Gabor filters to color
images using a specific form of hypercomplex numbers known as quaternions. These filters are
constructed as windowed basis functions of the quaternion Fourier transform also known as
hypercomplex Fourier transform. Based on this extension this paper presents the use of these new
quaternionic Gabor filters in colour texture image segmentation. Experimental results on two colour
texture images are presented. We tested the robustness of this technique for segmentation by adding
Gaussian noise to the texture images. Experimental results indicate that the proposed method gives better
segmentation results even in the presence of strongest noise.
Keywords
Colour texture image segmentation, Gabor filters, hypercomplex numbers, quaternions, quaternion
Fourier transform.
1. Introduction
Texture is a basic cue for human beings to recognize objects. Research on texture is a very
important task in computer vision and its applications. It has been a very active topic in the past
three decades. There are several research focuses in the field of texture analysis, mainly
including texture classification, texture segmentation, texture synthesis, shape from texture, etc
.Texture segmentation deals with textured image. The task is to segment a given image into
uniformly textured regions. This texture segmentation problem is one branch of the general
problem of image segmentation which is one important step in many computer vision tasks.
Regarding global variations of gray values or mean gray values over some neighborhood is in
most cases not sufficient for a correct segmentation.
The posed problem is rather vague since the term texture is not well defined and there is no
unique way of characterizing mathematically the local gray-value variations perceived as texture
by human observers. For this reason very different approaches to texture segmentation have
been taken. As local measure for the characterization of texture local statistical properties and
local geometric building blocks have been used. Another whole branch in texture segmentation
research is based on the local spatial frequency for characterizing texture. Gabor filters play a
special role in the analysis of local frequency. On the one hand the Gabor filter based
approaches to texture analysis are motivated by psychophysical research since 2D Gabor filters
have proven to be a good model for the cortical receptive field profiles while on the other hand
they are supported by the observation that a whole class of textures give rise to periodic gray
value structures[1].

Signal & Image Processing : An International Journal(SIPIJ) Vol.1, No.2, December 2010
76
In this paper we will restrict ourselves to the Gabor filter based approaches and introduce
quaternionic Gabor filter based approach to colour texture segmentation. Experimental results
have demonstrated the performance of this method. The remainder of this paper is organized as
follows. Section 2 briefly discusses the major related work. Section 3 presents the mathematics
of quaternions and Section 4 discusses representation of color pixels using quaternions. Section
5 speaks about quaternion Fourier transforms. Section 6 deals with quaternionic Gabor filters
and proposed texture segmentation algorithm. Section 7 presents the experimental results of
segmentation algorithm on colour texture images. Section 8 concludes the paper.
2. Related Work
There has been a lot of activity in texture segmentation using Gabor filters. We will comment on
some of these approaches. In [2, 3] Dunn introduces a method for finding an optimal 2D Gabor
filter for the discrimination of a two-texture image. This filter is designed such that the
discontinuity in the magnitude of the filtered image is most significant at the texture boundary.
In [4] Teuner et al. point out that the dominant frequencies do not necessarily coincide with the
ones that are important for segmentation. A dominant frequency is only helpful for
segmentation when it does not occur everywhere in the textured image. The authors provide a
measure called the spectral feature contrast which indicates the discriminatory power of Gabor
feature.
Bülow in [1] formulates quaternionic Gabor filters (QGF) by applying a Gaussian window to a
quaternionic Fourier transform and propose a method for texture segmentation of grayscale
images based on quaternionic Gabor filters. However, Bülow does not apply the quaternionic
Gabor filters to color or vector-valued images. Lilong Shi and Brian Funt in [5] propose a colour
texture segmentation method that makes possible encoding the structural and colour feature as a
whole unit by using the quaternion represented colour and use quaternion principal component
analysis(QPCA) with colours encoded as quaternions to calculate a basis for colour texture. Our
approach to colour texture segmentation is directly based on Bülow’s method. Here we extend
the Gabor filters to color images using quaternions and use these new quaternionic Gabor filters
in colour texture image segmentation.
3. Quaternions
The concept of the quaternion was introduced by Hamilton in 1843 [6]. It is the generalization
of a complex number. A complex number has two components: the real and the imaginary part.
The quaternion, however, has four components, i.e., one real part and three imaginary parts and
can be represented in Cartesian form as:

(1)
where , , and are real numbers and , and
are complex operators which obey the
following rules.

,
,

,

,

,

and also satisfies

1 .From these rules, it is clear that multiplication is
not commutative. The quaternion conjugate is

and the modulus of a
quaternion is given by
|| (2)

Signal Image Processing : An International Journal(SIPIJ) Vol.1, No.2, December 2010
77
A quaternion with zero real part is called a pure quaternion and a quaternion with unit modulus
is called a unit quaternion. The imaginary part of a quaternion has three components and may be
associated with a 3-space vector. For this reason, it is sometimes useful to consider the
quaternion as composed of a vector part and a scalar part. Thus q can be expressed as
, (3)
where the scalar part , is the real part i.e. , and the vector part is a composite of
three imaginary components,

(4)
Euler’s formula for the complex exponential generalizes to hypercomplex form

cosβ µ sinβ (5)
where µ is a unit pure quaternion. Any quaternion may be represented in polar form as
||µ
(6)
where µ and are referred to as the eigenaxis and eigenangle of the quaternion, respectively. µ
identifies the direction in 3-space of the vector part and may be regarded as a true generalization
of the complex operator , since µ

1. is analogous to the argument of a complex
number, but is unique only in the range [0 π] because a value greater than π can be reduced to
this range by negating or reversing the eigenaxis.
We can visualize the eigenaxis as the imaginary axis of an Argand diagram, the real axis of
which is aligned with the scalar axis of the quaternion 4-space. The eigenaxis is perpendicular to
the real axis, but need not be aligned with any of the three imaginary axes defined by the
imaginary operators , and
.
4. Quaternion Representation of Color Image Pixels
Color image pixels have three components, and they can be represented in quaternion form
using pure quaternions [7]. For images in RGB colour space, the three imaginary parts of a pure
quaternion can be used to represent the red, green and blue colour components. For example, a
pixel at image coordinates (x, y) in an RGB image can be represented as
!, , #, $,
(7)
where , ,#, and $, are the red, green and blue components of the pixel,
respectively.Fig.1 illustrates the approach adopted to represent RGB colour image using the
quaternion form.
Using quaternions to represent the RGB color space, the three color channels are processed
equally in operations such as multiplication. The advantage of using quaternion based
operations to manipulate color information in an image is that we do not have to process each
color channel independently, but rather, treat each color triple as a whole unit. We believe, by
using quaternion operations, higher color information accuracy can be achieved because a color
is treated as an entity.

78
Figure 1. Quaternion representation of a colour image.
5. Quaternion Fourier Transform
Based on the concept of quaternion multiplication and exponential, the Quaternion Fourier
Transform (QFT) has been introduced. Due to the non commutative multiplication rule of
quaternion algebra, there are several forms of quaternion Fourier tranforms.We adopted the
form presented in the work of [7][8] which divides the discrete QFT into two categories, namely
the right- side form and the left-side form.
Discrete version of the right-side and left- side quaternion Fourier transforms can be represented
as
Similarly, the inverse quaternion Fourier transforms can be denoted as:
In this transform, the hypercomplex operator was generalized: µ is any unit pure quaternion. µ
determines a direction in color space and an obvious choice for color images is the direction
corresponding to the luminance axis which connects all the points r=g=b. In RGB color space
this is the “gray line” and µ would be .

79
6. Quaternionic Gabor Filters
6.1 Definition
Based on existing work in monochrome applications of Gabor operations for texture
segmentation, we suggest that extending the Gabor techniques to color images will provide
good performance in analyzing texture images. To define a Gabor filter for color images, we
will include the concept of quaternions, a four-component hypercomplex number defined by Sir
William Hamilton [6].
A two-dimensional complex-valued Gabor filter is a linear shift-invariant filter with the impulse
response
%, #′
, ′
'exp 2,-. /. (12)
with
#, 01 2
1
2
34

56
7

4

5
7

89
where the coordinates ′
, ′
' are derived from , by a rotation about the origin through the
angle :.
4
,′
′
7 ;
cos α sin α

sin : cos :
= ;

=
and we will choose normalization constant K such that 0
6
?@?@
.The impulse response of
this Gabor filter is a Gaussian-windowed basis functions of the 2-D Fourier transform.
Bülow and Sommer in[1][9] formulate quaternionic Gabor filters (QGF) by applying a Gaussian
window to a quaternionic Fourier transform. However, they do not apply the quaternionic Gabor
filters to color or vector-valued images. To extend this form to quaternions, we first define a
pure unit quaternion A A6 A AB
with A. 0) and use it in place of the simple
imaginary root i. We then have
%, #, exp A2,-. 2,/. (13)
where #, 01 D
6

E;
F
?G
=

;
H
?@
=

IJ and 0
6
?@?@
with the result now a quaternion.
For a unit pure quaternion, Euler’s identity yields the relation K
cos L A MN L. Therefore,
the impulse response of a quaternionic Gabor filter is given by
%, #, OPM2,-. 2,/. A MN2,-. 2,/. (14)
and is a Gaussian windowed basis functions of the QFT. This is the filter that we will apply to
the colour texture segmentation. In the literature [7] the unit pure quaternion µ is stated as
arbitrary, and is set to the “gray axis” in RGB space – the resulting quaternion is
/
√3. A typical quaternionic Gabor filter is shown in Fig.2.
When performing a quaternionic Gabor filtering on the computer we have to use discrete
quaternionic Gabor filter masks of the form: % T%U,VWU,VXY6,…[ with
%U,V exp]
;U^
_`G
@
=
@
?G
@
;V^
_`G
@
=
@
?@
@ a b exp cA

[
4- ;d
[^6

= / ;N
[^6

=7e (15)

80
Using this convention the Gabor filter mask is an f b f quaternion matrix. The origin is
located at the center of the matrix, therefore it is advantageous to choose f odd, in order to
have a center pixel in the filter mask. The frequencies - and / count how many periods fit into
the filter mask in horizontal and vertical direction respectively. Fig. 3 shows the magnitude of
each component of a typical quaternionic Gabor filter of size 21b 21and for a single angle and
scale combination. The components show the usual Gabor profile of a sine wave attenuated by a
negative exponential. The different components have magnitudes related by the components of
the color axis vector µ.
(a) (b) (c) (d)
Figure 2. The four components of a quaternionic Gabor filter with parameters 56=3,5 3,
-. 0.1, /. 0. 1 and µ =
/√3 . The size of the filter mask is 21b 21. (a) real
component (b) i-component (c) j-component (d) k-component.
(a) (b)
(c) (d)

81
Figure 3. Mesh plots of Quaternionic Gabor Filter components with parameters 56=3,5
3, -. 0.1, /. 0. 1 and µ =
/√3 . The size of the filter mask is 21b 21. (a) real
component (b) i-component (c) j-component (d) k-component.
6.2 Local Quaternionic Phase
The local quaternionic phase of an image can be defined as the angular phase of the response to
a quaternionic Gabor filter. Each quaternion q given in Cartesian representation

can be represented in the form.
||hi
jk
lK
with (i, L, m)X[,, ,[ b[,/2, ,/2 Tb T,/4, ,/4W (16)
The i -and L components of the quaternionic phase correspond to the horizontal and vertical
Fourier phase while the m-component represents a new entity, which is not effected by a mere
shift of the image. The quaternionic phase angle m can be evaluated uniquely within the interval
T,/4, ,/4W as [1].
m =
oO MN2
/2 (17)
with || 1. The additional phase value m resulting from the quaternionic Gabor filtering
allows the distinction of the patterns in an image.
6.3 Texture Segmentation algorithm
Input: Texture image of size Nb N.
Output: Texture segmented image.
Step 1. Select values for -. , /. , 56 and 5 .
Step 2. Convert the input image into a 2-D quaternion matrix, say ,
Step 3. Obtain the impulse response of the filter, (which is also a 2-D quaternion matrix) by the
following formula.
%, #, exp A2,-. 2,/.
where #, 01 D
6

E;
F
?G
=

;
H
?@
=

IJ and 0
6
?@?@
Step 4. Compute the convolution of , with %, , say it is d, ,which is the
2-D quaternion matrix corresponding to the output image
d, = , p %,
Step 5. Obtain the output image. If it cleanly discriminates the textures then go to step 6, else
go to step 1 to select other values of -. , /. , 56 , 5 and repeat the procedure.
Step 6. Transform the filtered image into amplitude / phase representation.
Step 7. Extract the m -component of the local phase by applying the following formula to each
quaternion valued pixel
in the filtered image
m =
oO MN2
/2 with || 1
Step 8. Obtain the output segmented texture image by applying thresholding to the smoothed
m
component of the local quaternionic phase.
Step 9. Obtain the edges of the segmented image by applying a Sobel filter.
The steps involved in texture segmentation are shown in Fig. 4.

82
Texture
Image
Segmented
Texture image
Figure 4. Texture segmentation system.
7. Experimental Results
We demonstrate the segmentation process on two texture images as shown in Figs.5 (a) and
6(a). The input image is convolved with an optimally tuned quaternionic Gabor filter. We apply
one QGF whose central frequencies have been tuned to the main peak in the power spectrum
(QFT) of the image. The filtered image is transformed into amplitude / phase representation and
the component of the local phase is extracted. The component of the quaternionic phase
distinguishes not only local frequency and orientation but also local structure. The smoothed
component of the local quaternionic phase is thresholded. The edges of the
thresholded component are found by a Sobel filter. The results of segmentation are shown in
Figs.5 and 6. We tested the robustness of the for segmentation by adding Gaussian noise to
the textured image shown in Fig.5 (a). We added noise with zero mean and variance 0.2 and 0.4,
respectively. The results of segmentation with noise are shown in Figs.7 and 8. Although it is
almost impossible for a human observer to segment the image with the strongest noise, by
means of component most of the pixels are correctly classified.
(a) (b) (c)
(d) (e)
Figure 5. (a) The textured image (b) response of quaternionic Gabor filter mask of size 3 3
Demodulation
into amplitude
/phase form
Selection of filter
parameters
Convolution
with QGF
Extraction of
the -
component of
local phase
Thresholding
the -
component
Edge detection

83
(a) (b) (c)
(d) (e)
Figure 6. (a) The textured image (b) response of quaternionic Gabor filter mask of size 3 b 3
with parameters 56=3,5 3, -. 0.1 , /. 0. 1 and µ =
/√3 (c) the
m
component of the local quaternionic phase (d) the segmentation result after applying a
threshold (e) the edges of the m
component found by a Sobel filter.
(a) (b) (c)
(d) (e)

84
Figure 7. (a) The texture from figure 5(a) with added Gaussian noise of zero mean and variance
0.2 (b) the m
component of the local quaternionic phase (c) the median filtered m
component
(d) the segmented texture (e) the edges of the median filtered m
component found by a Sobel
filter showing the wrongly classified pixels.
(a) (b) (c)
(d) (e)
Figure 8. (a) The texture from figure 5(a) with added Gaussian noise of zero mean and variance
0.4 (b) the m
component of the local quaternionic phase (c) the median filtered m
component
(d) the segmented texture (e) the edges of the median filtered m
component found by a Sobel
filter showing the wrongly classified pixels.
8. Conclusion
Image segmentation is an essential step in pictorial pattern recognition and analysis applications.
Segmentation accuracy determines the success or failure of analysis procedures. Texture
segmentation is based on partitioning an image into different regions of similar textures based on
a specified criterion. In this paper, we presented a new colour texture image segmentation
method using quaternionic Gabor filters. On the basis of the QFT, we introduced quaternionic
Gabor filters for colour images. Application of these filters in colour texture segmentation has
been performed. In texture segmentation tasks the local quaternionic phase shows high
discriminatory power. It is an important fact that this additional feature comes without any
additional computational cost. We showed that the proposed texture segmentation method is
more robust against image noise. Two different textures, collected from standard album, are
used for experimentation. Future research will be concerned with extension of this application.
Acknowledgements
The authors are grateful to the Management, Principal and Head of the Department of
Electronics and Communication Engineering of Madanapalle Institute of TechnologyScience

85
for their constant support and encouragement. The authors also extend their gratitude to the
anonymous reviewers who have given very good suggestions for this better presentation of our
manuscript.
References
[1] T. Bülow. Hypercomplex Spectral Signal Representations for the Processing and Analysis of Images.
PhD thesis, Christian Albrechts University, 1999.
[2] D. Dunn and W.E. Higgins. Optimal Gabor filters for texture segmentation. IEEE Trans. Image
Processing, 4:947-964, 1995.
[3] D. Dunn, W.E. Higgins, and J.Wakeley. Texture segmentation using 2-d Gabor elementary functions.
T-PAMI, 16:130-149, 1994.
[4] A.Teuner, O. Pichler, and B.J. Hosticka. Unsupervised texture segmentation of images using tuned
matched Gabor filters. T-IP, 4:863-870, 1995.
[5] Lilong Shi,Brian Funt, “Quaternion color Texture Segmentation”,Computer Vision and image
understanding, Vol. 107, Issue 1-2,July,2007.
[6] W. R. Hamilton(1866). Elements of Quaternions. London, U.K: Longmans Green.
[7] Todd A. Ell and Stephen J. Sangwine “Hypercomplex Fourier Transforms of Color Images,” in IEEE
Transactions on Image Processing,Vol.16,No.1, pp. 22-35, 2007.
[8] T. A. Ell. “Quaternion Fourier transforms for analysis of 2-dimensional linear time-invariant partial-
differential systems,” in Proc. 32nd IEEE Conf. on Decision and Control, San Antonio, TX, pp.
1830-1841, 1993
[9] T. Bülow and G. Sommer. “Quaternionic Gabor filters for local structure classification,” in Proc. 14th
Annual Conf. on Pattern Recognition, Brisbane, Australia, pp. 808-810, 1998.
[10] Rafael C.Gonzalez, Richard E.Woods and StevenL.Eddins(2007). Digital Image Processing using
MATLAB: Pearson Education.
[11] S.Sangwine and N. Le Bihan, Quaternion Toolbox for Matlab, Software Library [Online]. Available:
http://qtfm.sourceforge.net.
[12] T. A. Ell. Hypercomplex Spectral Transformations. PhD thesis, University of Minnesota, 1992.
[13] Wang Hui; Wang Xiao-Hui; Zhou Yue; Yang Jie; Color Texture Segmentation Using Quaternion-
Gabor Filters”, Image Processing, 2006 IEEE International Conference on. 8-11 Oct. 2006
Page(s):745 – 748.
[14] Dawit Assefa,Lalu Mansinha, Kristy F. Tiampo, Henning Rasmussen and Kenzu Abdella; Local
quaternion Fourier transform and color image texture analysis Signal Processing, Vol. 90, Issue 6,
June 2010, Pages 1825-1835.
[15] Wang Xiao-Hui, Zhou Yue, Wang Yong-Gang and Zhu WeiWei; “Color Texture Segmentation
Based on Quaternion-Gabor Features” LNCS 4225, pp.345-353, 2006, Springer-Verlag Berlin
Heidelberg, 2006.
[16] Christoph Palm, Thomas M. Lehmann, Classification of color textures by Gabor filtering” Machine
Graphics Vision International Journal - Special issue on latest results in colour image processing
and applications archive ,Vol. 11, Issue 2/3, 2002.
[17] A. C. Bovik , M. Clark , W. S. Geisler, Multichannel Texture Analysis Using Localized Spatial
Filters, IEEE Transactions on Pattern Analysis and Machine Intelligence, Vol.12, no.1, pp.55-73,
January 1990.
[18] Anil K. Jain, Farshid Farrokhnia, Unsupervised texture segmentation using Gabor filters, Pattern
Recognition, Vol.24, no.12, pp.1167-1186, Dec. 1991.

86
[19] Yuzhong Wang, Jie Yang and Yue Zhou Color-texture segmentation using JSEG based on Gaussian
mixture modeling, Journal of Systems Engineering and Electronics, Vol. 17, Issue 1, March 2006,
Pages 24-29
Authors
B.D.Venkatramana Reddy is currently working as professor in ECE Department,
Madanapalle Institute of TechnologyScience, Madanapalle, India. He received his
M.Tech from S.V.University, Tirupathi, India. He has 12 years experience of teaching
undergraduate and post graduate students. He has published 12 research papers in
National/International conferences and journals. His research interests are in the areas of
signal processing and digital image processing.
Dr.T.Jayachandra Prasad obtained his B.Tech in Electronics and Communication
Engg., from JNTU College of Engineering, Anantapur 515002, and Master of
Engineering degree in Applied Electronics from Coimbatore Institute of Technology,
Coimbatore. He earned his Ph.D. Degree (Complex Signal Processing) in ECE from
JNTUCE, Anantapur, India. Dr.T.Jayachandra Prasad worked in KSRM College of
Engineering (KSRMCE), Kadapa, India from August 1984 to May 2006 in various
positions such as Assistant professor, Associate professor and Professor and HOD. He
worked as Head of ECE Dept. for 9 years at KSRMCE, Kadapa. He was instrumental
for the establishment of various laboratories at KSRMCE. Later he joined in RGM College of
Engineering and Technology, Nandyal, Kurnool (dt), Andhra Pradesh (state), INDIA. Presently, he is the
Principal of RGM College of Engineering and Technology, Nandyal. Dr.T.Jayachandra is having more
than 24 years of experience and has more than 18 technical publications in International journals and
National Journals. He is a life member of ISTE (India), Fellow of Institution of Engineers (Kolkata),
Fellow of IETE, Member of MIEEE and life member of NAFEN. His areas of interest include Signal
Processing and Image Processing.

Colour-Texture Image Segmentation using Hypercomplex Gabor Analysis

More Related Content

What's hot (20)

Similar to Colour-Texture Image Segmentation using Hypercomplex Gabor Analysis (20)

Recently uploaded (20)

Colour-Texture Image Segmentation using Hypercomplex Gabor Analysis