Fig. 1. The Radon transform g(s, ) is the 1-D projection of
f (x, y) at an angle .
Fig. 2. The Radon transform of a triangle shape (a) and the
Radon transforms of its translated (b), scaled (c), and rotated (d)
versions.
where s"x cos #y sin and u"! x sin #y cos
[25]. The Radon transform is linear, space-limited, and
periodic in with the period of 2. It has useful proper-
ties about translation, rotation, and scaling as outlined in
Eqs. (10)}(12).
translation: f (x!x
, y!y
)
0 g(s!x
cos !y
sin , ), (10)
rotation by
: f
N
(r, #
) 0 g(s, #
), (11)
scaling: f (ax, ay) 0
1
a
g(as, ), aO0. (12)
Here, f
N
(r, ) is the polar coordinate representation of
f (x, y) and the symbol `0a denotes the one-to-one trans-
formation relation. As shown in Eqs. (10)}(12), a transla-
tion of f (x, y) results in the shift of g(s, ) by a distance
equal to the projection of the translation vector (x
, y
)
on the line s"x cos #y sin . A rotation of the object
by angle
leads to a translation of its Radon transform
in the variable . A scaling of the (x, y) coordinates of
f (x, y) results in scaling of the s coordinate together with
an amplitude scaling of g(s, ). Fig. 2 shows the examples
of the properties of the Radon transform under di!erent
transformations.
3.1.2. HOS invariants generation scheme
Based on the properties of the Radon transform and
bispectra, we propose the following feature generation
method. For each model image pattern I(x, y), the orig-
inal 2-D data are "rst reduced to a set of 1-D functions
g(s, ) via the Radon transform using Eq. (9). It is as-
sumed that the origin is set at the center of the image
plane. Bispectrum of g(s, ) for each , denoted as
B
F
( f
, f
), is computed using
B
F
( f
, f
)"G
F
( f
)G
F
( f
)G
H
F
( f
#f
), (13)
where G
F
( f ) is the Fourier transform of g(s, ) with s be-
ing a variable and being a parameter. B
F
( f
, f
)is
translation invariant. A rotation of the object in the
image results in the cyclically shifted B
F
( f
, f
) along the
axis . The bispectrum of a scaled version of the original
image maintains the scaling e!ect except an additional
constant coe$cient in the magnitude of B
F
( f
, f
). The
bispectral moment M of order (p, q) of each B
F
( f
, f
)at
a certain angle is de"ned as
M
NO
()"
f
N
f
O
B
F
( f
, f
)df
d f
,
0)f
)f
)f
#f
)1, (14)
where p and q are two nonnegative integers. A set of
feature parameters
NO
() are then de"ned from the
bispectral moments M
NO
()as
NO
()"Imaginary(M
NO
() )/Real(M
NO
() ). (15)
In the following sections, we will prove that the de"ned
feature parameters
NO
() are RTS invariant.
2100 Y. Shao, M. Celenk / Pattern Recognition 34 (2001) 2097}2113