image description with charlier and meixner moments · image description with nonseparable...
TRANSCRIPT
IMAGE DESCRIPTION WITH
NONSEPARABLE TWO-DIMENSIONAL
CHARLIER AND MEIXNER MOMENTS
HONGQING ZHU*, MIN LIUy and YU LIz
Department of Electronics and Communications Engineering
East China University of Science and TechnologyShanghai 200237, P. R. China
*[email protected]@[email protected]
HUAZHONG SHUx and HUI ZHANG{
Laboratory of Image Science and TechnologySchool of Computer Science and Engineering
Southeast University
Nanjing 210096, P. R. [email protected]
This paper presents two new sets of nonseparable discrete orthogonal Charlier and Meixnermoments describing the images with noise and that are noise-free. The basis functions used by
the proposed nonseparable moments are bivariate Charlier or Meixner polynomials introduced
by Tratnik et al. This study discusses the computational aspects of discrete orthogonal Charlier
and Meixner polynomials, including the recurrence relations with respect to variable x andorder n. The purpose is to avoid large variation in the dynamic range of polynomial values for
higher order moments. The implementation of nonseparable Charlier andMeixner moments does
not involve any numerical approximation, since the basis function of the proposed moments isorthogonal in the image coordinate space. The performances of Charlier andMeixnermoments in
describing images were investigated in terms of the image reconstruction error, and the results of
the experiments on the noise sensitivity are given.
Keywords : Bivariate discrete orthogonal polynomials; nonseparable discrete orthogonal
moments; Charlier; Meixner; three-term recurrence relations; second order linear partialdi®erence equations.
1. Introduction
Moments have been extensively used in image processing, pattern recognition and
computer vision.1,8,13,19,24,25,28 Teague proposed continuous orthogonal polynomials
‡Author for correspondence
International Journal of Pattern Recognitionand Arti¯cial Intelligence
Vol. 25, No. 1 (2011) 37�55
#.c World Scienti¯c Publishing Company
DOI: 10.1142/S0218001411008506
37
as the basis functions to calculate continuous moments.16 These orthogonal poly-
nomials include Zernike polynomials, Legendre polynomials, and pseudo-Zernike
polynomials. Practical implementations of continuous orthogonal moments in digital
image analysis involve two main dilemmas9: (i) discrete approximation of the con-
tinuous integrals, and (ii) transformation of the image coordinate system into the
domain of polynomials. Consequentially, the development of moments led to their
discrete versions. Recently, Mukundan et al. and Yap et al. de¯ned discrete
orthogonal moments based on the discrete Tchebichef and Krawtchouk orthogonal
polynomials respectively.10,22 For the discrete orthogonal moments to be of use in
image processing, it must be extensible to two dimensions. Two forms of two-
dimensional moments have been suggested: a separable form and a nonseparable
form. A standard method computing two-dimensional separable form of moments is
via the row-column decomposition, which takes advantage of the separability of basis
functions of the moments.10,22,23 That is, the two-dimensional moment is broken up
into two one-dimensional moments computed in row-column or column-row wise
format. The nonseparable form moments do not work the same way; however, it can
be written using bivariate polynomials as the basis function of moments. This
requires less computation than for the separable moments. To the best of our
knowledge until now, no nonseparable form of two-dimensional moments has been
de¯ned in pattern recognition community. This is mainly due to the fact that
bivariate orthogonal polynomials are too di±cult to use. The purpose of this paper is
to introduce bivariate discrete orthogonal polynomials as basis functions of the
moments, and to expect that they have potential applications in image processing
and pattern recognition.
Actually, bivariate orthogonal polynomials have been investigated by many
authors.2�5,14,15,20,21 Bivariate orthogonal polynomials have been studied as the
generalization of one variable's orthogonal polynomials.3 However, comparing to the
theory of one variable, the structure of bivariate discrete orthogonal polynomial is
much more complicated. In 1982, Kowalski introduced a matrix notation for poly-
nomials applied in multiple variables, and showed the matrix notation with the
properties similar to three-term recurrence relations.4,5 Xu also presented a second-
order di®erence equations and bivariate discrete orthogonal polynomials in Refs. 20
and 21. More recently, Rodal et al. made a systematic study of the orthogonal
polynomials solutions of a second order partial di®erence equation with hypergeo-
metric type on two variables in Refs. 14 and 15. Although bivariate orthogonal
polynomials play an important role in applications such as numerical analysis, optics,
and quantum mechanics,11,12,18 much less attention has been paid to the application
of bivariate orthogonal polynomials in image analysis.
In this study, we have introduced the bivariate Charlier and Meixner polynomials
derived from the methods proposed by Tratnik.17 Then, the nonseparable form
bivariate moments, Charlier and Meixner, are constructed using their orthogonal
polynomials, respectively. Due to the new sets of moments being orthogonal in the
discrete domain of the image coordinate space, the proposed moments completely
38 H. Zhu et al.
eliminate the two problems referred to above. The image can be reconstructed from a
su±ciently large number of computed moments. In order to reduce the numerical
errors caused in the computation of Charlier and Meixner moments, this study
presents the three-term recurrence relations about variable x and polynomials
order n. The simulated results demonstrate that the proposed two-dimensional
nonseparable form moments Charlier and Meixner have better image description
capabilities.
The organization of this paper is given as follows: in Sec. 2, we recall the de¯nition
of the classical discrete orthogonal polynomials of one variable. Section 3 shows two
recurrence relations about univariate Charlier and Meixner polynomials. It is fol-
lowed by bivariate discrete orthogonal polynomials in Sec. 4. In Sec. 5, we de¯ne the
nonseparable discrete orthogonal moments with bivariable orthogonal polynomials
as basis functions. Some experimental results are presented and discussed in Sec. 6.
Section 7 concludes the study.
2. The Classical Discrete Orthogonal Polynomials of One Variable
The classical discrete orthogonal polynomials of one variable can be de¯ned as the
polynomial solutions of the following di®erence equation11,12
�ðxÞ�rpnðxÞ þ �ðxÞ�pnðxÞ þ �npnðxÞ ¼ 0 ð1Þwhere �pnðxÞ ¼ pnðxþ 1Þ � pnðxÞ, rpnðxÞ ¼ pnðxÞ � pnðx� 1Þ denote the forward
and backward ¯nite di®erence operators, respectively. �ðxÞ and �ðxÞ are functions ofsecond and ¯rst degrees respectively, �n is an appropriate constant. The solution of
this partial di®erence equation can be expressed by Rodrigues formula as follows:
pnðxÞ ¼Bn
wðxÞ rn½wnðxÞ� ð2Þ
where wðxÞ is weight function, wnðxÞ in the case of n ¼ 0. Thus the polynomials
solutions of Eq. (1) are determined by Eq. (2) depending on the normalizing factors
Bn. For the backward di®erence operator O we have the property11,12
rnfðxÞ ¼Xnk¼0
n
k
� �ð�1Þkfðx� kÞ ð3Þ
Combining Rodrigues formula and Eq. (3) one can obtain an explicit expression for
the polynomials pnðxÞ.The classical orthogonal polynomials of one discrete variable satisfy the following
three-term recurrence relation
xpnðxÞ ¼ �npnþ1ðxÞ þ �npnðxÞ þ �npn�1ðxÞ ð4ÞThe polynomials pnðxÞ satisfy an orthogonality relation of the form
Xsx¼0
pnðxÞpmðxÞwðxÞ ¼ d2n � �mn; 0 � m; n � s ð5Þ
Image Description with Nonseparable Two-Dimensional Charlier and Meixner Moments 39
where d2n denotes the square of the norm of the corresponding orthogonal poly-
nomials, and �mn denotes the Dirac function. The scaled orthogonal polynomials can
be obtained by utilizing the square norm and weighted function,
~pnðxÞ ¼ pnðxÞffiffiffiffiffiffiffiffiffiffiffiwðxÞd2n
sn ¼ 0; 1; s ð6Þ
Therefore, the orthogonal property of normalized orthogonal polynomials of Eq. (5)
can be rewritten as
Xsx¼0
~pmðxÞ~pnðxÞ ¼ �mn 0 � m; n � s ð7Þ
For univariate Charlier and Meixner polynomials, Nikiforov et al.11 introduced
some basis information listed in Table 1. Their weight functions are de¯ned in the
following discrete domain
G ¼ fx j 0 � x � sg ð8Þ
3. Computational Aspects of Univariate Charlierand Meixner Polynomials
In the following two subsections, we have attempted to present two recurrence
relations to calculate univariate Charlier and Meixner polynomials.
Table 1. Data for univariate Charlier ca1n ðxÞ, and univariate
Meixner $ð�;�Þn ðxÞ polynomials, (a1 > 0 for Charlier, and
0 < � < 1 for Meixner).
pnðxÞ ca1n ðxÞ $ð�;�Þn ðxÞ
s 1 1�ðxÞ x x
�ðxÞ a1 � x ��� xð1� �Þ�n n nð1� �ÞBn 1
an1
1
�n
wnðxÞ e�a1axþn1
x!
�xþn�ðnþ � þ xÞ�ð�Þx!
d2n
n!
an1
n!ð�Þn�nð1� �Þ�
�n �a1�
�� 1
�n nþ a1 nþ �ðnþ �Þ1� �
�n �n nðn� 1þ �Þ�� 1
40 H. Zhu et al.
3.1. Recurrence relations with respect to n
(i) Univariate Charlier polynomials
The Charlier polynomials of one variable ca1n ðxÞ satisfy the following ¯rst partial
di®erence equation of the form
x�rca1n ðxÞ þ ða1 � xÞ�ca1n ðxÞ þ nca1n ðxÞ ¼ 0 ð9Þwhere a1 is restricted to a1 > 0.
The nth Charlier polynomial is also de¯ned by using hypergeometric function as
follows:
ca1n ðxÞ¼2F0ð�n;�x; ;�1=a1Þ ð10ÞThe normalized discrete Charlier polynomials satisfy the following orthogonal
condition:
X1x¼0
~c a1mðxÞ~c a1
n ðxÞ ¼ �mn m � 0; n � 0 ð11Þ
The discrete Charlier polynomials satisfy the following three-term recurrence
relations:
A~c a1n ðxÞ ¼ B �D~c a1
n�1ðxÞ þ C � E~c a1n�2ðxÞ ð12Þ
where the coe±cients A� E in Eq. (12) are listed in Table 2.
The zero-order and the ¯rst-order normalizedCharlier can be calculated as follows:
~c a10 ðxÞ ¼
ffiffiffiffiffiffiffiffiffiffiffiwðxÞd20
s¼
ffiffiffiffiffiffiffiffiffiffiffiffiffiffie���x
x!
r
~c a11 ðxÞ ¼ �� x
�
ffiffiffiffiffiffiffiffiffiffiffiwðxÞd21
s¼ �� x
�
ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffie���xþ1
x!
r ð13Þ
(ii) Univariate Meixner polynomials
Meixner polynomials of one variable $ð�;�Þn ðxÞ satisfy the following ¯rst partial
di®erence equation of the form
x�r$ ð�;uÞn ðxÞ þ ð��� xð1� �ÞÞ�$ ð�;uÞ
n ðxÞ þ nð1� �Þ$ ð�;uÞn ðxÞ ¼ 0 ð14Þ
where � and � are restricted to 0 < � < 1, and � > 0.
Table 2. Data for three-term recurrence relations of univariate Charlier, and univariate Meixner
polynomials (a1 > 0 for Charlier, and � > 0, 0 < � < 1 for Meixner).
~pnðxÞ A B C D E
~c a1n ðxÞ �a1 x� nþ 1� a1 n� 1
ffiffiffiffiffiffia1n
r ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffia21
nðn� 1Þ
s
~$ð�;�Þn ðxÞ �
�� 1x� x�� nþ 1� �nþ �� ��
1� �
ðn� 1Þðn� 2þ �Þ1� �
ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi�
nð� þ n� 1Þr ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
�2
nðn� 1Þð� þ n� 2Þð� þ n� 1Þ
s
Image Description with Nonseparable Two-Dimensional Charlier and Meixner Moments 41
The nth Meixner polynomial is also de¯ned by using hypergeometric function as
follows:
$ ð�;uÞn ðxÞ ¼ ð�Þn2F1ð�n;�x;�; 1� 1=�Þ ð15Þ
The discrete normalized Meixner polynomials satisfy the following orthogonal
polynomials condition:
X1x¼0
~$ ð�;uÞm ðxÞ ~$ ð�;uÞ
n ðxÞ ¼ �mn m � 0; n � 0 ð16Þ
The discrete Meixner polynomials satisfy the following three-term recurrence
relations:
A ~$ ð�;�Þn ðxÞ ¼ B �D ~$
ð�;�Þn�1 ðxÞ þ C � E ~$
ð�;�Þn�2 ðxÞ ð17Þ
where the coe±cients A� E in Eq. (17) are listed in Table 2. The zero-order and the
¯rst-order normalized Meixner can be calculated as follows:
~$ð�;�Þ0 ðxÞ ¼
ffiffiffiffiffiffiffiffiffiffiffiwðxÞd20
s¼
ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi�xð� þ x� 1Þ!
x!ð� � 1Þ! ð1� �Þ�s
~$ð�;�Þ1 ðxÞ ¼ � þ x� x
�
� � ffiffiffiffiffiffiffiffiffiffiffiwðxÞd21
s¼ � þ x� x
�
� � ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi�xð� þ x� 1Þ!
x!ð� � 1Þ!�ð1� �Þ�
�
s
ð18Þ
According to Table 1, the above recurrence relations Eqs. (12) and (17) can
be used to calculate the weighted univariate Charlier and Meixner polynomials,
respectively.
3.2. Recurrence relations with respect to x
Mukundan discussed the reasons in Ref. 7 that the recursive nature of polynomial
evaluation can lead to numerical problems when the required moment order is large.
An e®ective solution to this problem is changing the recurrence relation to avoid
cumulative multiplication of large values. In this subsection, we present x recurrence
relations to eliminate the numerical instability while computing polynomials or high-
order moments. The general frame of their recurrence relations with respect to x is as
follows:
Considering the properties of the operators O and 4, we have
4OpnðxÞ ¼ pnðxþ 1Þ � 2pnðxÞ þ pnðx� 1Þ ð19Þ
42 H. Zhu et al.
Thus, the recursive relation of discrete orthogonal polynomials with respect to x can
be obtained according to Eqs. (1) and (19) as follows:
pnðxÞ ¼2�ðx� 1Þ þ �ðx� 1Þ � �n
�ðx� 1Þ þ �ðx� 1Þ pnðx� 1Þ � �ðx� 1Þ�ðx� 1Þ þ �ðx� 1Þ pnðx� 2Þ ð20Þ
The recurrence relation of normalized discrete orthogonal polynomials is also written
using the following general form:
~pnðxÞ ¼ffiffiffiffiffiffiffiffiffiffiwðxÞp
�ðx� 1Þ þ �ðx� 1Þ
� 2�ðx� 1Þ þ �ðx� 1Þ � �nffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiwðx� 1Þp ~pnðx� 1Þ � �ðx� 1Þffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
wðx� 2Þp ~pnðx� 2Þ" #
ð21Þ
Thus, one can obtain the recurrence relations of discrete Charlier, and Meixner
with respect to x according to Eq. (21) and Table 1 as follows:
~c a1n ðxÞ ¼
ffiffiffiffiffiffiffiffiffiffiwðxÞp
�ðx� 1Þ þ �ðx� 1Þ
� 2�ðx� 1Þ þ �ðx� 1Þ � �nffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiwðx� 1Þp ~c a1
n ðx� 1Þ � �ðx� 1Þffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiwðx� 2Þp ~c a1
n ðx� 2Þ" #
ð22Þ
~$ ð�;�Þn ðxÞ ¼
ffiffiffiffiffiffiffiffiffiffiwðxÞp
�ðx� 1Þ þ �ðx� 1Þ
� 2�ðx� 1Þ þ �ðx� 1Þ � �nffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiwðx� 1Þp ~$ ð�;�Þ
n ðx� 1Þ � �ðx� 1Þffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiwðx� 2Þp ~$ ð�;�Þ
n ðx� 2Þ" #
ð23ÞUsing Eqs. (2), (3), (6) and Table 1, we obtain the initial values of recurrence
relations with respect to x listed in Table 3. The above equations can be used to
e®ectively calculate the weighted univariate Charlier and Meixner polynomials'
values. Figure 1 shows the plots of the ¯rst few orders of these polynomials with
di®erent parameter values.
Table 3. The initial values of three-term recurrence relations with
respect to x for univariate Charlier and Meixner polynomials. (a1 > 0for Charlier, and 0 < � < 1; � > 0 for Meixner).
~pnðxÞ x ¼ 0 x ¼ 1
~c a1n ðxÞ ffiffiffiffiffiffiffiffiffiffi
wð0Þd2n
sa1 � n
a1
ffiffiffiffiffiffiffiffiffiffiwð1Þwð0Þ
s~c a1n ð0Þ
~$ð�;�Þn ðxÞ ð�Þn
ffiffiffiffiffiffiffiffiffiffiwð0Þd2n
s�ðnþ �Þ � n
��
ffiffiffiffiffiffiffiffiffiffiwð1Þwð0Þ
s~$ ð�;�Þ
n ð0Þ
Image Description with Nonseparable Two-Dimensional Charlier and Meixner Moments 43
4. Bivariate Discrete Orthogonal Polynomials
4.1. Bivariate Charlier polynomials
The bivariate Charlier polynomials were introduced by Tiatnik17 and Rodal
et al.14,15 For a bivariable, let n1;n2 2 N0, one has
ca1;a2n1;n2ðx; yÞ ¼ ~$ ð�x�y;�a1=a2Þ
n1ðxÞ~c ða1þa2Þ
n2ðxþ y� n1Þ ð24Þ
where ~$ð�;�Þn ðxÞ is the normalized univariateMeixner polynomials de¯ned by Eq. (23),
and ~c anðxÞ is the normalized univariate Charlier polynomials de¯ned by Eq. (22).
Charlier polynomials of two variables satisfy the following orthogonality relation,
X1x¼0
X1y¼0
ca1;a2n1;n2ðx;yÞca1;a2m1;m2
ðx;yÞ ax1a
y2
x!y!e�a1�a2
¼ n1!n2!a2a1
� �n1
ða1 þ a2Þn1�n2�n1;m1�n2;m2
ð25Þ
0 50 100 150 200 250−0.2
−0.15
−0.1
−0.05
0
0.05
0.1
0.15
0.2
x
Poly
nom
ials
Val
ues
Charlier Polynomials (a1 = 128, Order:10 )
(a)
0 50 100 150 200 250−0.15
−0.1
−0.05
0
0.05
0.1
0.15
0.2
x
Poly
nom
ials
Val
ues
Charlier Polynomials (a1 = 240, Order:10 )
(b)
0 50 100 150 200 250−0.2
−0.15
−0.1
−0.05
0
0.05
0.1
0.15
0.2
x
Poly
nom
ials
Val
ues
Meixner Polynomials (β = 60, µ = 0.5, Order:10 )
(c)
0 50 100 150 200 250−0.2
−0.15
−0.1
−0.05
0
0.05
0.1
0.15
0.2
x
Poly
nom
ials
Val
ues
Meixner Polynomials (β = 150, µ = 0.5, Order:10 )
(d)
Fig. 1. Plots of univariate polynomials when N ¼ 256, Order ¼ 10, (a) Charlier (a1 ¼ 128), (b) Charlier
(a1 ¼ 240), (c) Meixner (� ¼ 60; � ¼ 0:5), and (d) Meixner (� ¼ 150; � ¼ 0:5).
44 H. Zhu et al.
The bivariate polynomials uðx; yÞ ¼ ca1;a2n;m ðx; yÞ satisfy the following di®erence
equation
x�1r1uðx; yÞ þ y�2r2uðx; yÞ þ ða1 � xÞ�1uðx; yÞþ ða2 � yÞ�2uðx; yÞ þ ðnþmÞuðx; yÞ ¼ 0 ð26Þ
Figure 2 shows the plots of bivariate polynomials along y direction, the para-
meters a1 and a2 vary from 20 to 50.
4.2. Bivariate Meixner polynomials
The multivariate Meixner polynomials were introduced by Tiatnik17 and Rodal
et al.14,15 For a bivariable, let n1;n2 2 N0, we have
$�;�1;�2n1;n2
ðx; yÞ ¼ ~$ ð�x�y;��1=�2Þn1
ðxÞ ~$ ð�þn1;�1þ�2Þn2
ðxþ y� n1Þ ð27Þ
0 10 20 30 40 50 60−0.06
−0.04
−0.02
0
0.02
0.04
0.06
Bivariate Charlier Polynomials (a1= 20, a
2= 50, Order= 6 )
y
u(x,
y)
(a)
0 10 20 30 40 50 60−0.06
−0.04
−0.02
0
0.02
0.04
0.06
0.08
Bivariate Charlier Polynomials (a1= 25, a
2= 25, Order= 6 )
y
u(x,
y)(b)
0 10 20 30 40 50 60−0.06
−0.04
−0.02
0
0.02
0.04
0.06
Bivariate Charlier Polynomials (a1= 50, a
2= 20, Order= 6 )
y
u(x,
y)
(c)
Fig. 2. Plots of bivativate Charlier polynomials uðx; yÞ ¼ ca1 ;a2n;m ðx; yÞ along y direction, when N ¼ 60,
Order ¼ 4, (a) a1 ¼ 20, a2 ¼ 50, (b) a1 ¼ 25, a2 ¼ 25 and (c) a1 ¼ 50, a2 ¼ 20.
Image Description with Nonseparable Two-Dimensional Charlier and Meixner Moments 45
where �1 and �2 are nonzero real parameters satisfying
j�1j þ j�2j < 1 ð28Þand ~$
ð�;�Þn ðxÞ is the univariate Meixner polynomials de¯ned by Eq. (23), orthogonal
with respect to
wðxÞ ¼ �xð�Þx�ðxþ 1Þ ; � > 0 and 0 < � < 1 ð29Þ
Meixner orthogonal polynomials uðx; yÞ ¼ $�;�1;�2n1;n2
ðx; yÞ of two variables satisfy
the following orthogonal relation
X1x¼0
X1y¼0
$�;�1;�2n1;n2
ðx; yÞ$�;�1;�2m1;m2
ðx; yÞ �x1�
y2ð�Þxþy
x!y!ð1� �1 � �2Þ�
¼ ��;�1;u2n1;n2
�n1;m1�n2;m2
ð30Þ
where,
��;�1;�2n1;n2
¼ n1!n2!ð�Þn1þn2ð�1 þ �2Þn1�n2
�2
�1ð1� �1 � �2Þ� �
n1
ð31Þ
The bivariate Meixner polynomials satisfy the following di®erence equation
ð1� �2Þx�1r1uðx; yÞ þ ð1� �1Þy�2r2uðx; yÞþ �1y�1�2uðx; yÞ þ �2x�2r1uðx; yÞþ ðð�1 þ �2 � 1Þxþ ��1Þ�1uðx; yÞþ ðð�1 þ �2 � 1Þyþ ��2Þ�2uðx; yÞ� ð�1 þ �2 � 1Þðn1 þ n2Þuðx; yÞ ¼ 0 ð32Þ
Figure 3 shows the plots for the ¯rst few orders of the bivariate Meixner polynomials
with the parameters �, �1 and �2 varying from 0.15 to 0.35. The bivariate Charlier
and Meixner polynomials with di®erent parameters are listed in Table 4.
5. Two-Dimensional Discrete Orthogonal Moments
Given a digital image fðx; yÞ with size N �N , i.e. x 2 ½0;N � 1� and y 2 ½0;N � 1�,the (mþ n)th order moments with bivariate orthogonal polynomials as basis func-
tion of an image is de¯ned as follows
Mmn ¼XN�1
x¼0
XN�1
y¼0
fðx; yÞuðx; yÞ ð33Þ
Here, bivariate polynomials uðx; yÞ ¼ $�;�1;�2n1;n2
ðx; yÞ and uðx; yÞ ¼ ca1;a2n;m ðx; yÞ are
used to construct the nonseparable Charlier and nonseparable Meixner moments.
46 H. Zhu et al.
Using the orthogonal property of normalized orthogonal polynomials, Eq. (33)
also leads to the following inverse moment transform.
fðx; yÞ ¼Xsm¼0
Xs
n¼0
Mmnuðx; yÞ ð34Þ
where s > 0 for bivariate Charlier and Meixner polynomials. It indicates that an
image can be completely reconstructed by calculating its discrete orthogonal
moments. If moments are limited to an order K, we can approximate f by f̂ .
f̂ ðx; yÞ ¼XKm¼0
XK�m
n¼0
Mmnuðx; yÞ ð35Þ
0 10 20 30 40 50 60−0.06
−0.04
−0.02
0
0.02
0.04
0.06
0.08Bivariate Meixner Polynomials (β = 60, µ
1= 0.15, µ
2= 0.35, Order:6 )
y
u(x,
y)
(a)
0 10 20 30 40 50 60−0.06
−0.04
−0.02
0
0.02
0.04
0.06
0.08Bivariate Meixner Polynomials (β = 60, µ
1= 0.25, µ
2= 0.25, Order:6 )
y
u(x,
y)(b)
0 10 20 30 40 50 60−0.06
−0.04
−0.02
0
0.02
0.04
0.06
0.08Bivariate Meixner Polynomials (β = 60, µ
1= 0.35, µ
2= 0.15, Order:6 )
y
u(x,
y)
(c)
Fig. 3. Plots of bivariate Meixner polynomials uðx; yÞ ¼ $�;�1 ;�2n1 ;n2
ðx; yÞ along y direction, when N ¼ 60,
Order ¼ 6, � ¼ 60, (a) �1 ¼ 0:15, �2 ¼ 0:35, (b) �1 ¼ 0:25, �2 ¼ 0:25, and (c) �1 ¼ 0:35, �2 ¼ 0:15.
Image Description with Nonseparable Two-Dimensional Charlier and Meixner Moments 47
6. Experimental Results
In this section, three experiments have been carried out to verify the image rep-
resentation capability of the proposed moments when they are used for gray-level
image or binary image. The test gray-level images of the size 128� 128 pixels used
for this study are shown in Fig. 4. The di®erence between an image and a version
reconstructed from a ¯nite set of its moments is a good measure of the image rep-
resentation capability of the set of moments considered. Hence, the ¯rst experiment
reconstructed the images by the moments. The di®erence between the original image
and the reconstructed image is measured using the mean squared error (mse). The
Table 4. Bivariate orthogonal polynomials with di®erent parameters.
Bivariate Charlier polynomials
a1= 15, a
2= 15 a
1= 15, a
2= 30 a
1= 15, a
2= 60
a1= 30, a
2= 15 a
1= 30, a
2= 30 a
1= 30, a
2= 60
a1= 60, a
2= 15 a
1= 60, a
2= 30 a
1= 60, a
2= 60
µ1= 0.15, µ
2= 0.15 µ
1= 0.15, µ
2= 0.25 µ
1= 0.15, µ
2= 0.35
µ1= 0.25, µ
2= 0.15 µ
1= 0.25, µ
2= 0.25 µ
1= 0.25, µ
2= 0.35
µ1= 0.35, µ
2= 0.15 µ
1= 0.35, µ
2= 0.25 µ
1= 0.35, µ
2= 0.35
Bivariate Meixner polynomials ( =50) β
(a) (b) (c)
Fig. 4. Test images (128� 128): (a) Lena, (b) Flower, (c) Girl.
48 H. Zhu et al.
mean squared error is de¯ned as follows:
mse ¼ 1
N 2
XN�1
x¼0
XN�1
y¼0
½fðx; yÞ � f̂ ðx; yÞ�2 ð36Þ
where fðx; yÞ and f̂ ðx; yÞ denote the original image and the reconstructed image,
respectively.
The reconstructed images by the nonseparable Charlier and Meixner moments are
shown in Fig. 5. In order to set a global feature extraction model and obtain better
reconstruction results, we set � ¼ 70, �1 ¼ �2 ¼ 0:25 for nonseparable Meixner
moments, and a1 ¼ a2 ¼ 60 for nonseparable Charlier moments. In this case, non-
separable Charlier and Meixner polynomials shall emphasize the central region of the
image. Due to the in°uence of parameters on orthogonal polynomials, the global and
local information of an image can be emphasized by choosing the right values of the
parameters. In our previous study, we had discussed how to obtain a global feature
extraction model by choosing appropriate parameters.26,27 Owing to the limit of the
space, this study omits the discussion of choosing proper parameters for the proposed
nonseparable moments. From Fig. 5, it can be seen that the images reconstructed by
Order=30 Order=60 Order=90 Order=128
Fig. 5. Reconstruction of the gray-level image Flower and Girl. The order's numbers from left to right are
30, 60, 90, and 128, respectively. The ¯rst two rows are reconstructed images using nonseparable Meixner
moments (� ¼ 70; �1 ¼ 0:25; �2 ¼ 0:25). The last two rows are reconstructed images using nonseparableCharlier moments (a1 ¼ a2 ¼ 60).
Image Description with Nonseparable Two-Dimensional Charlier and Meixner Moments 49
the nonseparable Charlier are clearly of better quality than those by nonseparable
Meixner. Figure 6 shows the plotting of mse values for nonseparable Charlier and
Meixner moments in a maximum order up to 128. The reconstruction errors decrease
monotonically with the increase of the order's values as predicted. It also demon-
strates that the images reconstructed by the nonseparable Charlier moments are
much more like the original images with the increasing of moment's order. From
these results, one can conclude that moments constructed by bivariate Charlier
polynomials have better global extraction capability than the moments constructed
by bivariate Meixner polynomials.
The aims of the second experiment are to compare the performances of the pro-
posed nonseparable Charlier and Meixner moments with that of the Legendre
moments16 and Tchebichef moments.10 Image reconstruction processes are repeated
for the gray-level image Lena shown in Fig. 7. The images' reconstruction errors mse
by using di®erent methods are shown in Fig. 8. Both Figs. 7 and 8 show that
0 20 40 60 80 100 1200
0.05
0.1
0.15
0.2
0.25
0.3
0.35
Moment Order
Rec
onst
ruct
ion
Err
or
110 115 120 125 130
1.5
2
2.5
3
3.5
4
x 103
Bivariate Meixner MomentsBivariate Charlier Moments
(a)
0 20 40 60 80 100 1200
0.05
0.1
0.15
0.2
0.25
Moment Order
Rec
onst
ruct
ion
Err
or
110 115 120 125 130
2
3
4
5
6
7x 10
3
Bivariate Meixner MomentsBivariate Charlier Moments
(b)
Fig. 6. Comparative study of reconstruction errors by using nonseparable Charlier moments
(a1 ¼ a2 ¼ 60), and nonseparable Meixner moments (� ¼ 70; �1 ¼ 0:25; �2 ¼ 0:25). (a) Image Flower,
(b) image Girl.
(a) (b) (c) (d)
Fig. 7. Reconstruction of the gray-level image Lena, order is 128, (a) Legendre moments, (b) Tchebichefmoments, (c) nonseparable Charlier moments (a1 ¼ a2 ¼ 60), (d) nonseparable Meixner moments
(� ¼ 70; �1 ¼ 0:25; �2 ¼ 0:25).
50 H. Zhu et al.
Legendre moments give the worst results comparing to both Tchebichef moments
and the proposed moments. One also observes that the reconstruction results based
on nonseparable Charlier and Meixner moments are not as good as Tchebichef
moments.
Sensitivity to noise is a critical issue for image moments. Only when the values of
the moments are not sensitive to noise can they be ideal features of an image. In the
¯nal experiment, we have investigated and compared the behaviors of the proposed
moments in noisy images. The discussion will concentrate on binary images Hen and
Snow°ower which are used in the MPEG-7 Core Experiment CE-Shape-1 part B.6
(This database consists of 70 shape categories with 20 objects per category.) How-
ever, the extension to gray-level images is straightforward. Table 5 presents the
0 20 40 60 80 100 1200
0.05
0.1
0.15
0.2
0.25
Moment Order
Rec
onst
ruct
ion
Err
or
110 115 120 125 130
4
6
8
10
12
14
16x 103
LegendreTchebichefCharlierMeixner
Fig. 8. MSE comparison of images reconstructed using Legendre, Tchebichef, nonseparable Charlier
(a1 ¼ a2 ¼ 60), and nonseparable Meixner moments (� ¼ 70; �1 ¼ 0:25; �2 ¼ 0:25), respectively.
Table 5. Reconstructed images of noisy images with Legendre, Tchebichef, Charlier (a1 ¼ a2 ¼ 60)
and Meixner moments (� ¼ 70; �1 ¼ 0:25; �2 ¼ 0:25). 5% salt-and-pepper noises for image Hen, �2 ¼0:1 Gaussian white noises for image Snow°ower.
Original Images Noisy Images Legendre Tchebichef Charlier Meixner
Image Description with Nonseparable Two-Dimensional Charlier and Meixner Moments 51
reconstructed results for noisy images, Hen and Snow°ower, sized 128� 128. The
¯rst column represents the original binary images. The second column shows two
noisy images. Here, binary image Hen is corrupted by 5% salt-and-pepper noises, and
�2 ¼ 0:1 Gaussian white noises are added in image Snow°ower. Figure 9 shows the
mse as a function of the total number of moments used in the reconstruction. From
Fig. 9, one can see that the reconstruction error decreases monotonically with the
increase of the moments' orders. The images reconstructed from Tchebichef and
nonseparable Charlier moments have much better quality than those from the
nonseparable Meixner and Legendre moments. The nonseparable Meixner moments
also yield lower reconstruction error than Legendre moments. These experimental
results are approximately consistent with those obtained by the second experiment.
7. Conclusions
Although moment functions had been extensively utilized for a number of image
processing tasks, no bivariate discrete orthogonal polynomials had been used as basis
function sets to directly de¯ne the two-dimensional moments. In this paper, we
introduced two nonseparate discrete orthogonal moments: Charlier and Meixner
moments. The image representation capability of the proposed moments had been
veri¯ed. We did the comparison of two most recent orthogonal moments: continuous
orthogonal Legendre moments and discrete orthogonal Tchebichef moments. Ex-
perimental results showed that the proposed moments were comparable to Legendre
moments and moderately poorer than Tchebichef moments.
The main contribution of this study is as follows: (1) This work has presented two
formulae to obtain x and n based recurrence relations calculating the univariate
Charlier and Meixner polynomials. (2) We de¯ned two-dimensional nonseparable
discrete orthogonal moments based on bivariate discrete orthogonal polynomials in
0 20 40 60 80 100 1200
0.05
0.1
0.15
0.2
0.25
0.3
0.35
Moment Order
Rec
onst
ruct
ion
Err
or
110 115 120 125 130
0.035
0.04
0.045
0.05
0.055
0.06
0.065
0.07LegendreTchebichefCharlierMeixner
(a)
0 20 40 60 80 100 1200
0.05
0.1
0.15
0.2
0.25
0.3
0.35
Moment Order
Rec
onst
ruct
ion
Err
or
110 115 120 125 130
0.03
0.04
0.05
0.06
0.07
0.08
0.09
0.1LegendreTchebichefCharlierMeixner
(b)
Fig. 9. Reconstruction errors for noisy images using Legendre, Tchebichef, Charlier (a1 ¼ a2 ¼ 60), and
Meixner moments (� ¼ 70; �1 ¼ 0:25; �2 ¼ 0:25), respectively. (a) Image Hen with 5% salt-and-pepper
noises; (b) image Snow°ower with �2 ¼ 0:1 Gaussian white noises.
52 H. Zhu et al.
the image processing and pattern recognition ¯eld. (3) The proposed method here
could be easily extended to obtain other bivariate orthogonal polynomials and their
corresponding moments.
Due to the nonsepariability of bivariate orthogonal polynomials, Charlier and
Meixner moments' values could not be computed in two steps by successive one-
dimensional operations on rows and columns of an image. Thus, pixel by pixel
calculation was needed to obtain the moments' values, and this process was very
expensive. Further on, our work will focus on the fast algorithm to calculate
nonseparable Charlier or Meixner moments.
Acknowledgments
The authors would like to thank the anonymous referees for their helpful comments
and suggestions. This work has been supported by National Natural Science Foun-
dation of China under Grant No. 60975004.
References
1. X. B. Dai, H. Z. Shu, L. M. Luo, G. N. Han and J. L. Coatrieux, Reconstruction oftomographic images from limited range projections using discrete Radon transform andTchebichef moments, Patt. Recogn. 43 (2010) 1152�1164.
2. C. F. Dunkl and Y. Xu, Orthogonal Polynomials of Several Variables, Encyclopedia ofMathematics and Its Applications (Cambridge University Press, Cambridge, 2001).
3. T. Koornwinder, Two-variable Analogues of the Classical Orthogonal Polynomials, inTheory and Application of Special Functions (Academic Press, 1975).
4. M. A. Kowalski, The recursion formulas for orthogonal polynomials in n variables, SIAMJ. Math. Anal. 13 (1982) 309�315.
5. M. A. Kowalski, Orthogonality and recursion formulas for polynomials in n variables,SIAM J. Math. Anal. 13 (1982) 316�376.
6. MPEG7 CE Shape-1 Part B: http://www.imageprocessingplace.com/root ¯les V3/image databases.htm
7. R. Mukundan, Some computational aspects of discrete orthogonal moments, IEEETrans. Image Process. 13 (2004) 1055�1059.
8. R. Mukundan, Fast computation of geometric moments and invariants using Schlick'sapproximation, Int. J. Patt. Recogn. Artif. Intell. 22 (2008) 1363�1377.
9. R. Mukundan, S. H. Ong and P. A. Lee, Discrete vs. continuous orthogonal moments forimage analysis, Int. Conf. Imaging Science, Systems and Technology-CISST'01 (2001),pp. 23�29.
10. R. Mukundan, S. H. Ong and P. A. Lee, Image analysis by Tchebichef moments, IEEETrans. Imag. Process. 10 (2001) 1357�1364.
11. A. F. Nikiforov and V. B. Uvarov, Special Functions of Mathematical Physics (Boston,Bessel, 1988).
12. A. F. Nikiforov, S. K. Suslov and V. B. Uvarov, Classical Orthogonal Polynomials of aDiscrete Variable (Springer, New York, 1991).
13. G. A. Papakostas, E. G. Karakasis and D. E. Koulouriotis, Novel moment invariants forimproved classi¯cation performance in computer vision applications, Patt. Recogn. 43(2010) 58�68.
Image Description with Nonseparable Two-Dimensional Charlier and Meixner Moments 53
14. J. Rodal, I. Area and E. Gogoy, Orthogonal polynomials of two discrete variables on thesimplex, Integral Transf. and Spec. Func. 16 (2005) 263�280.
15. J. Rodal, I. Area and E. Gogoy, Linear partial di®erence equations of hypergeometrictype: Orthogonal polynomials solutions in two discrete variables, J. Comput. Appl. Math.200 (2007) 722�748.
16. M. R. Teague, Image analysis via the general theory of moments, J. Opt. Soc. Amer. 70(1980) 920�930.
17. M. V. Tratnik, Some multivariable orthogonal polynomials of the Askey tableau-discretefamilies, J. Math. Phys. 32 (1991) 2337�2342.
18. A. Vercin, Ordered products, W1-algebra, and two-variable, de¯nite-parity, orthogonalpolynomials, J. Math. Phys. 39 (1998) 2418�2427.
19. B. Wang and Y. Q. Chen, An invariant shape representation: Interior angle chain, Int. J.Patt. Recogn. Artif. Intell. 21 (2007) 543�559.
20. Y. Xu, On discrete orthogonal polynomials of several variables, Adv. Appl. Math. 33(2004) 615�632.
21. Y. Xu, Second order di®erence equations and discrete orthogonal polynomials of twovariables, Int. Math. Res. Notices. 8 (2005) 449�475.
22. P.-T. Yap, R. Paramesran and S.-H. Ong, Image analysis by Krawtchouk moments,IEEE Trans. Image Process. 12 (2003) 1367�1377.
23. P.-T. Yap, R. Paramesran and S.-H. Ong, Image analysis using Hahn moments, IEEETrans. Patt. Anal. Mach. Intell. 29 (2007) 2057�2062.
24. P.-T. Yap and P. Raveendran, An e±cient method for the computation of Legendremoments, IEEE Trans. Patt. Anal. Mach. Intell. 27 (2005) 1996�2002.
25. F. Zhang, S.-Q. Liu, D.-B. Wang and W. Guan, Aircraft recognition in infrared imageusing wavelet moment invariants, Imag. Vis. Comput. 27 (2009) 313�318.
26. H. Zhu, H. Shu, J. Zhou, L. Luo and J. L. Coatrieux, Image analysis by discreteorthogonal dual-Hahn moments, Patt. Recogn. Lett. 28 (2007) 1688�1704.
27. H. Zhu, H. Shu, J. Liang, L. Luo and J. L. Coatrieux, Image analysis by discreteorthogonal Racah moments, Sign. Process. 87 (2007) 687�708.
28. J. Žunić, K. Hirota and P. L. Rosin, A Hu moment invariant as a shape circularitymeasure, Patt. Recogn. 43 (2010) 47�57.
54 H. Zhu et al.
Hongqing Zhu obtainedher Ph.D. in 2000 fromShanghai Jiao Tong Uni-versity. From 2003 to2005, she was a post-doctoral fellow at theDepartment of Biologyand Medical Engineering,Southeast University.Currently she is anassociate professor at East
China University of Science & Technology.Her current research interests mainly focus
on image reconstruction, image segmentation,image compression, and pattern recognition.
Min Liu received theM.Sc. degree in data tele-communications and net-works from University ofSalford, U.K. in 2001.After carrying outresearch, in 2005 sheobtained the Ph.D. degreein wireless laser com-munications. Currentlyshe is a lecturer at the
East China University of Science and Technol-ogy (ECUST), China.
Her research interests include optical com-munications and pattern recognition.
Yu Li received his Ph.D.degree from ZhejiangUniversity in 2007. In2007, he joined EastChina University of Sci-ence and Technology.Now, he is a teacher in theDepartment of Electronicand CommunicationEngineering.
His research interestsinclude image signal processing, array signalprocessing, etc. He has published 11 papersincluding nine journal papers. He is a member ofIEICE.
Huazhong Shu (M'00-SM'06) received the B.S.degree in applied math-ematics from WuhanUniversity, China, in1987, and a Ph.D. innumerical analysis fromthe University of Rennes(France) in 1992. He isnow with the Departmentof Computer Science and
Engineering of Southeast University, China.His recent work has focused on image anal-
ysis, pattern recognition and fast algorithms ofdigital signal processing. Dr. Shu is a seniormember of IEEE.
Hui Zhang received theB.S. degree in radioengineering in 2003 andM.S. degree in biomedicalengineering in 2006 fromSoutheast University, re-spectively. He is currentlypursuing the Ph.D. at theDepartment of ComputerScience of Southeast Uni-versity.
His research interests are mainly focused onpattern recognition, image watermarking, imageregistration and image processing.
Image Description with Nonseparable Two-Dimensional Charlier and Meixner Moments 55