1 2004
mplate
Hence
which
isation
ts and
> done
erence
1dence
iching,
'emote
South
g and
quares
gs Of
atation
digital
Yodels,
emote
ptimal
ethod,
rained
/ and
iturity,
demote
19.
OSILIOR
greek)
ns. 24
"y and
d UK,
ing of
nmeltry
3. 254-
ecision
«of
= Vol.
itching
fo
Swiss
Vision,
SAR IMAGE SEGMENTATION THROUGH B-SPLINE DEFORMABLE CONTOURS AND
FRACTAL DIMENSION
J. Gambini *, M. Mejail *, J. Jacobo-Berlles ", C. Delrieux RE
* Universidad de Buenos Aires, FCEyN, Departamento de Computación
Ciudad Universitaria, Pab. I, 1428 Buenos Aires, Repüblica Argentina
Tel/FAX: 454 11 4576-3359, e-mail: (jgambini, marta, jacobo } @dc.uba.ar
** Universidad Nacional del Sur, Departamento de Ingenieria Eléctrica
Bahia Blanca, Republica Argentina Tel. and FAX: ++54 291 4595154
e-mail: claudio@acm.org
Working Group WG III/4
KEY WORDS: change detection, classification, edge extraction, feature segmentation, SAR imagery
ABSTRACT
Synthetic Aperture Radar (SAR) images are usually co
This makes difficult the segmentation, object identificati
we propose the combination of local fractal estimation an
extraction problem in SAR images. After a supervised initi
pletely within the region of interest), the algorit
boundary of the region to be segmented. Th
in the surrounding. Fractal dimension provi
measurement of the fractal dimension is widely ackno
putational requirements. Box counting algorithms are
pixels in a surrounding of varying sizes.
of pixels with a given brightness profile, an
formed. The proposed algorithm is systematically tested on synthetic and rea
performance of our proposal are assessed.
1 INTRODUCTION
Segmentation of region boundaries is an important issue in
remote sensing image analysis and many techniques have
been studied to solve this problem. In the particular case
of region boundary determination in SAR images, the use
of Active Contours has been developed by O. Germain et
al. (Germain, 2001), J. Jacobo et al. (Jacobo-Berlles et al.,
2002) and M. J. Gambini ef al. (Gambini et al., 2004) in
all cases with a statistical approach.
The technique proposed in this work is based on B-Spline
boundary fitting and was originally developed by A. Blake
et al. (Blake and Isard, 1998). Blake’s proposal cannot
be successfully applied to SAR images due to the speckle
noise, since local brightness fluctuations induce spurious
control points, and therefore a bad boundary segmentation.
Here, we adapt the active contours idea, but the control
points are chosen in places where a local gradient in the
fractal dimension are found. The boundary extraction pro-
cess we describe here, begins with the application of local
fractal dimension (box counting). Histogram classification
of this attribute of the image allows an adequate thresh-
old detection. Then, the determination of zones of interest,
each of them includes one region for which we are going
to extract the corresponding boundary. These zones of in-
terest are given by B-Spline curves, determined by their
1159
rrupted by a signal-dependent non-additive noise called speckle.
on, and feature extraction within this kind of images. In this work
d B-Spline based active contours as a solution for the boundary
alization (the specification of a an initial curve laying com-
hm searches the control points (vertices) of a B-Spline curve that fits the
e vertices of the curve are found by a local estimation of the fractal dimension
des a good local roughens and statistical correlation estimation. Box-counting
wledged to be the most adequate in terms of robustness and com-
based on a statistical analysis of the brightness distribution of the
A power law can be established between the surrounding size and the amount
d then an adequate assessment of the local fractal dimension can be per-
| SAR images, and both the accuracy and the
control points. Then, a series of segments is drawn on the
image, and the image data lying on them are extracted. For
each segment, the transition point, that is, the point be-
longing to the region’s boundary, is determined. This de-
termination is based on the estimated values for the fractal
dimension. Then, for each region, the contour sought is
given by the B-spline curve that fits these transition points.
The structure of this paper is as follows: section 2 de-
scribes the model used here for fractal dimension estima-
tion, section 3 gives an introduction to B-Spline curve fit-
ting, section 4 specifies the criterion used to determine the
transition points and explains the region fitting algorithm,
section 5 shows the obtained results and finally, section 6
presents the conclusion an further research .
2 FRACTAL ANALYSIS FOR IMAGE
SEGMENTATION
Fractal dimension analysis has been largely used as an im-
age descriptor for segmentation. The strategies that we
present here seem to produce better results in autonomous
classification problems that had not been addressed before.
A local image estimator that can be useful when coping
with noisy images is the fractal dimension (Mandelbrot,
