ACTIVE CONTOUR MODEL TO DETECT LINEAR FEATURES
IN SATELLITE IMAGES
M. R. Della Rocca ^', M. Fiani *, A. Fortunato *, P.Pistillo *
* Dipartimento di Ingegneria Civile, Università di Salerno, 84084 Fisciano, Italy —
(r.dellarocca, m.fiani, a.fortunato, ppistill)@unisa.it
KEY WORDS: Remote Sensing, Extraction, Detection, Algorithms, Features, Satellite, Segmentation.
ABSTRACT:
Active contour models are used extensively in image processing applications, including edge detection, shape modelling, segmenta-
tion and in general to detect object boundaries. They are curves defined in the image domain that can move under the influence of in-
ternal forces within the curve itself and external forces derived from the image data. Many versions and improvements have been
proposed since then, in an effort to improve on several aspects of the original finite difference implementation.
We present preliminary results of a wavelet based active contour model for detecting linear features in satellite images The method
combines multi-scale decomposition and edge detection in a fast-converging iterative scheme. As an application, coastline retrieval
from a SAR image is illustrated.
I. INTRODUCTION
Satellite images can hardly be automatically segmented, due to
their inherent complexity. Low signal-to-noise ratio, undesired
images features and other factors further complicate this issue.
Manual and semi-automatic tracking of images is still the
mainstream method for obtaining good segmentations of
complicated images. Many algorithms have been proposed to
extract features from satellite images.
The subject of the present study is a group of high level
segmentation models, the so called active contour models for
detecting linear features.
Active contour models, also known as snakes, are used extensi-
vely in image processing applications, including edge detection,
shape modeling, medical image-analysis, to detect object boun-
daries. Snakes are curves defined in the image domain that can
move under the influence of internal forces within the curve i-
tself and external forces derived from the image data. The inter-
nal and external forces are defined so that the snake will even-
tually conform to an object boundary or some other desired
image feature.
Problems associated with initialization and poor convergence to
concave boundaries, however, have hitherto limited their use.
They were first proposed in 1987 by Kass, Witkin and Terzo-
poulos. Many versions and improvements have been proposed
since then, in an effort to improve on several aspects of the ori-
ginal finite difference implementation.
In fact the original snake model presents a number of limita-
tions. First, the initial contour should be sufficiently close to the
object, to prevent converge to wrong results. Second, the per-
formance of the snake depends on the number of control points,
which is usually fixed. Besides, the method is unable to extract
the multiple-objects contours and runs into difficulties when fa-
cing concave boundaries.
In this paper we apply an algorithm which uses a modified ver-
sion of the active countour model which features a new class of
external forces to addresses the problems listed above.
Generally, the most common method used to detect an edge
contour in an image is to set the external energies as the negati-
ve modulus of the gray level gradient of the image.
We define this energy as the negative of the modulus of a wave-
let transform of the image. In particular we utilize wavelet tran-
sforms to obtain a filtered image at a certain scale. The desired
contour is accordingly identified, through the active contour
model, working on.the filtered image. Then, the obtained
contour is taken as the initial position of the snake on a
wavelet-filtered image corresponding to a more accurate scale.
The process ends when the scale (detail) level of the original
image has been reached.
The algorithm has been applied to identify different features
and detect linear characteristics in satellite images.
2. SNAKE MODEL
A snake (Kass et al., 1987) is a parametric curve defined in the
image domain which is initialized manually by a set of control
points, lying on an open or closed curve.
v(s) = (x(s), y(5)) se [0,1] (1)
Associated to a snake is an energy function which is used to
move the snake across the image. For each control point, the
energy is recalculated for all points in its neighborhood and the
point that minimizes this energy function is used to update the
control point. Once the update procedure settles, one has
hopefully detected a feature of interest (edge), which can be
reconstructed by interpolation among the control points.
The energy functional to be minimized is defined as
E uem. HE) Eus (0())) ds @)
0
, Corresponding author. r.dellarocca@unisa.it; tel. and fax: 39 89 964366.
446
Internat
where /
where |
snake's
denote t
s.
The sec
forms of
In (4),
intensit!
A snake
which c
The int
the ext
image c
To find
V as fu
A solut
the velc
vanishe
In the
represe
scheme
Severa
atal, |
et al,
analyti
which
Variou
