3D meas-
lerivation
rrain sur-
ce Model
cesses on
areas or
ficult the
fast and
informa-
ation for
nally the
acknow-
jally Dr.
ôffler are
die Laser-
ft für Ver-
1. [1992],
j für die
ungsuesen
ne und di-
eodätische
tric Week
bration of
Heipke &
hotogram-
ational So-
;, Munich,
near Edge
S.
Fusion for
ited at the
Jonference
June 1996.
ot Vision,
Laser Sys-
Data Ac-
ional Soci-
omo, Italy,
pographis-
eodätische
r Scanner-
7.
SHAPE DISCRIMINATION BY DESCRIPTORS AND MOMENTS USING NEURAL NETWORK
Kyoung-Ok Kim*, Young-Kyu Yang*, Yong-Hui Park**, Tae-Kyun Kim**
Eoeun-Dong, Yusung-ku, Taejeon, Korea 305-333
*Systems Engineering Research Institute/ Korea Institute of Science and Technology
**Dept. of Computer Eng., Chungnam National Univ. Goong-Dong, Yusung-Ku, Taejeon, Korea
Tel: *82-42- 869-1401 Fax: *82-42- 869-1479
E-mail: ykyang@seri.rekr, kokimQGseri.re.kr
Commission III, Working Group 3
KEY WORDS : Vision, Training, Feature, Neural
ABSTRACT
An important problem in target recognition is the automatic discrimination of the object in a scene
regardless of its position, size, and orientation. Object recognition is processed by feature extraction and
similarity measurement. This paper is to recognize target objects using moments and Fourier descriptors.
The Fourier descriptors and moment features are used as input vectors to the neural network classifier.
The difference between the features is that the former deals with contours, while the latter deals with
area. This paper presents preprocessing technique and the performance comparison of Zernike moment,
Hu's moment invariant and Fourier descriptors as features. Noise is another important factor to affect
the recognition accuracy. The contour smoothing as preprocessing for Fourier descriptor is adopted for
noise removal.
INTRODUCTION
Several methods have been studied for object
recognition in computer vision and pattern
recognition. The process of feature extraction is a
very important step in object recognition.
The current approaches to invariant 2D shape
recognition include extraction of global image
information using regular moments, boundary-base
analysis via Fourier descriptors, autoregression
models, image representation by circular harmonic
expansion, syntactic, and neural network
approaches (Kotanzard, 1990) But the global
approaches doesn't work very well for occluded
objects, so local features should be considered for
partially occluded objects. Moments have been
utilized as object feature in a number of
application for this purpose. The Hu's seven
nonlinear functions (Hu, 1962) defined on regular
moments are one of the popular type of moments.
(Dudant, 1977) But the basis set is not
orthogonal. Recently Zernike moment which is
known to have strong class separability power is
getting popular to derive feature vectors. (Kim,
1994). Moreover Zernike moments used in this
International Archives of Photogrammetry and Remote Sensing. Vol. XXXI, Part B3. Vienna 1996
study are a class of orthogonal moments. Fourier
descriptors which are different from area based
moments extract contour features.
Advantages, disadvantages and performance
comparison to Fourier descriptors, Zernike
moments and moment invariants are also
discussed in this paper.
FEATURE VECTORS
Moments invariants
(p*q)th moment is defined as
wy, = Sly a) (1)
Central moments can be normalized to become
invariant to scale change by defining
=
E
um s. pere 1 (2)
r
ex
389
