Simple Image Analysis By Moments

Version 0.2∗
Maintained by: Johannes Kilian
March 15, 2001

Abstract
This document is a preliminary version of a manual on using moments to
analyze the form of objects in an image. Only moments up to second order
are used to get simple features of the object, like center of gravity, semi-axes,
orientation and eccentricity of the object ellipsis.
If you have comments or can contribute to this document, please mail J. Kilian

c
°This document is freely distributable.

Contents
1 General remarks on moments 1
1.1 Definition . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1
1.2 Order of moments . . . . . . . . . . . . . . . . . . . . . . . . . . 2
1.3 Spatial, central and central normalized moments . . . . . . . . . 3

2 Moments in OpenCV 4

3 Object Features based on moments 5

3.1 Area A of the object . . . . . . . . . . . . . . . . . . . . . . . . . 5
3.2 Center of Gravity xc , yc . . . . . . . . . . . . . . . . . . . . . . . 5
3.3 Tensor of moments . . . . . . . . . . . . . . . . . . . . . . . . . . 6
3.3.1 Semi-major and Semi-major axes a and b . . . . . . . . . 6
3.3.2 Orientation θ . . . . . . . . . . . . . . . . . . . . . . . . . 6
3.4 Roundness κ and Eccentricity ε . . . . . . . . . . . . . . . . . . . 7

1 General remarks on moments

1.1 Definition
The definition of moments of the gray value-function f (x, y) of an object is the
following:
∗ This document is for public distribution

1
 Z Z
mp,q = xp y q f (x, y)dxdy (1)

The integration is calculated over the area of the object. Generally each
other pixel based feature instead of the gray value could be used to calculate
the moments of the object.
Using binary images the gray value function f (x, y) becomes

1 Object
f (x, y) = b(x, y) = { (2)
0 Background
and can be neglected in the subsequent formulas.

1.2 Order of moments

Moments are generally classified by the order of the moments. The order of a
moment depends on the indices p and q of the moment mp,q and vice versa.
The sum p + q of the indices is the order of the moment mp,q .
Considering this, the following moments are defined:

• zero order moment ((p, q) = (0, 0))

Z Z
m0,0 = dxdyb(x, y) (3)

The zero order moment describes the area A of the object.

• first order moments ((p, q) = (1, 0) or (0, 1))

Z Z
m1,0 = dxdyxf (x, y)
Z Z
m0,1 = dxdyyf (x, y) (4)

The first order moments contain information about the center of gravity
of the object;
m1,0
xc =
m0,0
m0,1
yc = (5)
m0,0

• second order moments ((p, q) = (2, 0) or (0, 2) or (1, 1)))

Z Z
m2,0 = dxdyx2 f (x, y)
Z Z
m0,2 = dxdyy 2 f (x, y) (6)
Z Z
m1,1 = dxdyxyf (x, y)

• ...

2
1.3 Spatial, central and central normalized moments
Formula (1) describes general spatial moments of the object.
From the spatial moments the central moments can be derived by reducing
the spatial moments with the center of gravity (xc , yc ) of the object, so all the
the central moments refer to the center of gravity of the object. Expressed as
formula the central moments are calculated as follows
Z Z
µp,q = (x − xc )p (y − yc )q f (x, y)dxdy (7)

From this the following facts result:

µ0,0 = m0,0
µ1,0 = µ0,1 = 0 (8)

The central moments of first or higher order can directly be derived from
the spatial moments by
µ ¶p µ ¶q
mp,q m1,0 m0,1
µp,q = − ∗ (9)
m0,0 m0,0 m0,0
Using formula (9) the central moments of first and second order can be
derived from spatial moments as follows:
µ ¶
m1,0 m1,0
µ1,0 = − =0
m0,0 m0,0
µ ¶
m0,1 m0,1
µ0,1 = − =0
m0,0 m0,0
µ ¶2
m2,0 m1,0 m2,0
µ2,0 = − = − x2c (10)
m0,0 m0,0 m0,0
µ ¶2
m0,2 m0,1 m0,2
µ0,2 = − = − yc2
m0,0 m0,0 m0,0
µ ¶ µ ¶
m1,1 m1,0 m0,1 m1,1
µ1,1 = − ∗ = − xc ∗ yc
m0,0 m0,0 m0,0 m0,0
where xc , yc is the center of gravity of the object (see formula (5)).
The main advantage of central moments is their invariancy to translations
of the object. Therefore they are suited well to describe the form of the object.
A disadvantage of the spatial and central moments is their dependency on
the size of the object. This is disturbing, when trying to compare objects, which
are looked at from different distances. To allow a comparison of those objects,
scaling of the moments is necessary. Normally the area A of the object is used
as a scaling factor. Dividing the central moments µp,q with powers of A we get
central normalized moments νp,q .
µp,q
νp,q = p+q (11)
+1
m0,02
The main advantage of normalized moments is their invariancy to the size
of the object.

3
2 Moments in OpenCV
As a data structure to maintain the moments the moment state structure
CvMoments is defined in file CvTypes.h as follows:

typedef struct CvMoments {

double m00, m10, m01, m20, m11, m02, m30, m21, m12, m03;
double mu20, mu11, mu02, mu30, mu21, mu12, mu03;
double inv_sqrt_m00;
} CvMoments;

Description of the members of the CvMoments-structure:

m00 . . . m03 Spatial moments mp,q of zero to third order
mu20 . . . mu03 Central moments µp,q of second and third order. By definition
µ0,0 = m0,0 and µ1,0 = µ0,1 = 0, therefore only second and third order
central moments are stored in the structure.
inv sqrt m00 is the inverse square root of the spatial moment m00. This value
can be used to calculate the normalized moments νp,q from the central
moments µp,q
There are several functions in OpenCV, dealing with moments:

• cvMoments() - calculates moments up to third order of image plane and

fills moment state structure. (see OpenCV Reference Manual, Chapter 6
- Image Statistics).
• cvGetSpatialMoment() - retrieves spatial moment from the moment state
structure. (see OpenCV Reference Manual, Chapter 6 - Image Statistics).
• cvGetCentralMoment() - retrieves central moment from the moment state
structure. (see OpenCV Reference Manual, Chapter 6 - Image Statistics).
• cvGetNormalizedCentralMoment() - retrieves normalized central moment
from the moment state structure. (see OpenCV Reference Manual, Chap-
ter 6 - Image Statistics).
• cvGetHuMoments() - calculates seven moment invariants from the mo-
ment state structure. (see OpenCV Reference Manual, Chapter 6 - Image
Statistics).
• cvContoursMoments() - calculates contour moments up to order 3.(see
OpenCV Reference Manual, Chapter 3 - Contour Processing).

Information about moments are found in Chapter 3 (Contour Processing)

and Chapter 6 (Image Statistics) of the OpenCV Reference Manual.
As the calculation of moments in OpenCV with cvContourMoments()is
based on the contour processing of OpenCV, the moments can only be de-
termined on binary images. To use the contour moments to perform a simple
object analysis, the object has to be separated in the image: the considered
object has to be separated from the rest of the image with binarization.

4
 It has to be considered that the moments calculated in OpenCV do have
negative values, because cvContourMoments() is affected by the contour orien-
tation of OpenCV. Because contours are retrieved in mirrored coordinate system
(origin is top-left, y-axis is oriented downward), all the contours do have clock-
wise orientation and thus do have negative area according to Green’s formula.
From this fact result negative moments with clockwise oriented contours.
Errors in cvContourMoments(): Up to version Nov 08 2000 there are se-
vere errors in calculating the central moments, but the spatial moments are
calculated correctly. For this reason, use formula (9) to determine the central
moments correctly! Also the sign of the spatial moment is determined wrongly
(it is positive in each case and should be negative in some cases (with clock-
wise oriented contours)). This errors should be corrected in version later than
version Nov 08 2000

3 Object Features based on moments

The moments are features of the object, which allow a geometrical reconstruc-
tion of the object. They do not have a direct understandable geometrical mean-
ing, but usual geometrical parameters can be derived from them.
In the following sections only moments up to second order are used to derive
fairly simple object features. These features exactly describe simple objects up
to the complexity of an ellipsis.
To get more precise description of complex objects, you have to use moments
of higher order and more complex moments like Zernike- or Legendre-moments.
The more complex your object is, the higher the order of your describing mo-
ments should be to get a minimal error reconstructing your object by your
moments.

3.1 Area A of the object

Considering the gray value function f (x, y) as the density of the object ρ(x, y),
the spatial moment of zero order m0,0 is the area A of the object.

A = m0,0 (12)
OpenCV offers a separate function cvContourArea() to calculate the area
of the object bounded by the contour, but it could also retrieved from member
m00 of the struct CVMoments.

3.2 Center of Gravity xc , yc

As mentioned in formula (5) the coordinates xc and yc of the center of gravity
are simply described by the spatial moments of first order m1,0 and m0,1 divided
by the zero order moment m0,0 (i.e. the area A of the object).

m1,0 m1,0
xc = =
A m0,0
m0,1 m0,1
yc = = (13)
A m0,0

5
 y
y’

a b
x
Θ
x’

Figure 1: Image Ellipse

3.3 Tensor of moments

The analogy of the moments to mechanical moments allows a deeper under-
standing of the central moments of second order µ2,0 , µ0,2 and µ1,1 . They
contain terms, in which the gray value function f (x, y), i.e. the density ρ(x, y)
of the object is multiplied with the square of the distance from the center of
gravity (xc , yc ). Exactly the same terms are available in the inertial tensor,
known from physical mechanics. The three central moments of second order
build the components of the inertial tensor of the rotation of the object about
its center of gravity: · ¸
µ2,0 −µ1,1
J= (14)
−µ1,1 µ0,2
Using the inertial tensor analogy several further parameters could be derived
from the central moments of second order.

3.3.1 Semi-major and Semi-major axes a and b

The main inertial axis could be derived by calculating the eigenvalues of the
inertial tensor:
r q
1
λ1,2 = ∗ (µ2,0 + µ0,2 ) ± 4 ∗ µ21,1 − (µ2,0 − µ0,2 )2 (15)
2
The main inertial axes of the object correspond to the semi-major and semi-
minor axes a and b of the image ellipse which can be used as a approximation
of the considered object. The main inertial axes are those axes, around which
the object can be rotated with minimal (major semi-axis a) or maximal (minor
semi-axis b) inertia.
The semi-axis a and b are illustrated in figure (1).

3.3.2 Orientation θ
The orientation of the object is defined as the tilt angle between the x-axes and
the axis, around which the object can be rotated with minimal inertia (i.e. the
direction of the major semi-axis a). This corresponds to the eigenvector with
minimal eigenvalue. In this direction the object has its biggest extension. It is
calculated as follows:

6
 1 2µ1,1
θ= arctan (16)
2 µ2,0 − µ0,2
The tilt angle θ is illustrated in figure (1).
There is an ambiguity in the tilt angle θ of the object which can be resolved
by choosing θ always to be the angle between the x-axis and the semimajor axis
a, i.e. by definition a ≥ b
Secondly, we pick the principal value of the arc tangent such that − π2 ≤
arctan x ≤ π2
With this results for the tilt angle θ are arrived that are given in Table 1.

3.4 Roundness κ and Eccentricity ε

Further parameters to describe the form of the object are the roundness κ and
eccentricity ε. Both of give a measure for the roundness of the considered object.
The roundness κ can easily be calculated by dividing the square of the
perimeter p with the area A:

p2
κ= (17)
A
Because a circle has the maximal Area A within a given perimeter p, a
scaling of roundness κ is performed:

p2
κ= (18)
2πA
Therefore κ for a circle is equal 1, for other objects > 1.
The eccentricity ε can directly derived from the semi-major and semi-minor
axes a and b of the object:
√
a2 − b2
ε= (19)
a
ε can be directly calculated from the central moments of second order by

(µ2,0 − µ0,2 )2 − 4µ21,1

ε= (20)
(µ2,0 + µ0,2 )2

µ2,0 − µ0,2 µ1,1 θ

Zero Zero 0o
Zero Positive +45o
Zero Negative −45o
Positive Zero 0o
Negative Zero −90o
1 2µ1,1
Positive Positive 2 arctan µ2,0 −µ0,2 0o < θ < 45o
1 2µ1,1
Positive Negative 2 arctan µ2,0 −µ0,2 −45o < θ < 0o
1 2µ1,1
Negative Positive 2 arctan µ2,0 −µ0,2 + 90o 45o < θ < 90o
1 2µ1,1
Negative Negative 2 arctan µ2,0 −µ0,2 − 90o −90o < θ < −45o

Table 1: Tilt angle θ (orientation) of an image ellipsis

7
 The eccentricity ε can have values from 0 to 1. It is 0 with a perfectly round
object and 1 by a line shaped object.
The eccentricity ε is a better measure than the roundness κ of the object,
because it has a clearly defined range of values and therefore it can be compared
much better.

References
[Jähne, 1997] Jähne, B. (1997). Digitale Bildverarbeitung. Springer-Verlag,
Berlin-Heidelberg, 4. edition.
[Teague, 1980] Teague, M. R. (1980). Image analysis via the general theory of
moments. Journal of the Optical Society of America, 70(8):920–930.