| figures)
onsidera-
itivity is
he calcu-
this pro-
(b)
(c)
3.1.3 Affine-invariant Fourier descriptors
The points of the closed contour may be viewed
as being in a complex image plane. Starting at
an arbitrary point on the contour, and tracking
once around it, yield a sequence of complex num-
bers. The discrete Fourier transform of this se-
quence defines the so-called Fourier descriptors
of the contour. In this coefficients the frequencies
are represented. Experimentally we found that 16
coefficients are sufficient for a geometrically satis-
fying description of the simple geometric features.
To get affine-invariant features the complex Fou-
rier descriptors are aggregated in pairs. Each pair
defines a 2 x 2 matrix of which the determinant
is calculated. This yields scalar quantities which
are also called A-invariants (Arbter, 1989). The
A-invariants or more precisely the absolute A-
invariants found by a further normalization are
the primary quantities for the backprojection into
the spatial domain. By calculating the inverse
LE a CA
ame S
k=0.604 K-0.785 k=1.0
Figure 4: Starting situation (upper row), the nor-
med figures and the corresponding form factors
(lower row)
Fourier transform the normed figures plotted in
figure 4 are obtained. From these figures the form
factor k — 47 F/U? is determined which defines
the compactness of the object. The calculated
form factors (see figure 4) are a simple feature
for the characterization and identification of the
unknown object.
3.1.4 Affine-invariant moments
Given a two-dimensional function g(z, y) the cen-
tral moments of order p -- q can be expressed as
Hog = / Fi -— £)"(y — y)*g(z, y) dz dy
—OO —oO
where z,j are the weighted mean values. From
this moments up to order 3 four affine-invariant
features ],,152,13,14 can be derived. The main
part of the derivation is the evaluation of a qua-
dratic form. The resulting formula, for example
of I, reads as
I — (p2opoz — p31) / pao -
The formulas for the other quantities are more
extensive. The interested reader should consult
the paper by Hu (1962).
In contrast to the procedures described before the
determination of affine-invariant features using
the moments relies on the area within the border
line. Some experimental results for this invari-
ants are listed in table 1. In this case the binary
image of the region of interest defines the func-
tion g(z, y), i.e., all pixels within the border line
of the corresponding figure are equal to 1. The
I I, I; I,
triangle
t: 9.25.1079 -3.25-10-7 -5.49-10-5 4.06-10-°
e: 9.26.107? -3.23-10-7 -5.47-10-5 4.06-10-°
$s... 7.82038. ..2.1-10-8 2.0-107% 1.8107
rectangle
t: 6.94-10-3 0 0 0
e: 6.94.10? 0 0 0
s: 6.8.1078 0 0 0
circle
t: 6.25.1073 0 0 0
e: 6.34.107? .82.10-16 .2.1.107? 1.0.10-4
s 17110/* 73-10-16 1.0.107? 5.2.10-!!
Table 1: Invariants I;, I2, I3, I4 for triangle, rec-
tangle and circle. (t: theoretical, e: experimental,
s: standard deviation)
theoretical values are derived analytically. The
experimentally found mean values refer to a cer-
tain number of calculations with differently affine
distorted figures. The empirical values and the
theoretical values fit well together. The largest
differences can be observed for the circle. In ad-
dition to the mean values the corresponding stan-
dard deviations are listed. The calculated values
467