1 increasing attention
wo broad categories:
tion is presented. The
ion of the geometric
ING
rmed, it is necessary
There are two main
oise and distributed
e only in some pixel
may be termed as a
Iges is only local and
ll pixels and may be
therefore appropriate
e found between the
s filters and the noise
| smoothing, so non-
letection. The most
] for treating impulse
thing (EPS) and the
| images with a small
than EPS, since it
more accurate edge
otherwise. We used
g stage, setting its
n regions satisfying a
on texture or edge
used, based on two
direction. In order to
nd to ease the road
| be approximated by
Ist be introduced on
ong as the gradient
ant along contiguous
1 is a line segment.
y the gradient vector
old may be fixed for
yut therefore will be
, either making life
hem from getting any
age preprocessing on
until we are in the
condition to discard what becomes clearly useless.
In the gradient computation large masks tend to increase
smoothing, loosing details; small masks instead preserve
fine detail, but are very sensitive to noise. We used a small
2x2 mask (see Figure 1) which also gives an invariant
response with respect to line rotations (Burns et al., 1986).
m E
ee
Figure 1. The mask used for computing the gradient
For each pixel in the image, the gradient magnitude is
computed and, if its value is larger than the chosen
threshold, the orientation is computed as well. In the
following, when speaking of image or image orientation we
will always refer to this part of the original image. To
proceed with the feature extraction, all contiguous pixels
enjoying similar gradient orientation are grouped in regions,
because they are likely to belong to the same edge. The
space of the orientations [0-2x] is divided into suitable
equally spaced intervals, the so-called partitions (see Figure
2).
09
BUS ce
E vs xa5
mec
i /
BENE
225
180^
Figure 2. Gradient space partitioning in 8 intervals
2.3 Feature extraction
We look for a description of the image content based on
lines. This may be achieved in many ways, e.g. by line
following, relaxation, Hough transform etc. (Ballard &
Brown, 1982); we opted for an alternative suggested by
(Burns et al., 1986), with some minor changes. The concept
is the following: we get a line segment from each region
where the gradient orientation is in a certain range. The
straight line to which the segment belongs is defined by the
gravity centre of the area and by the direction perpendicular
to gradient direction. The end points of the segment are
determined by projecting the points of the area over the
straight line.
The segment orientation is computed by a robust method
(either Hampel, Huber or the L-1 norm may be selected),
which some experiment proved to be better than a total least
squares approach, particularly in small regions and with a
small number of partitions. The gravity centre is computed
as a weighted mean, using the gradient magnitude.
The choice of the number of partitions is critical: if there are
too many we get a very fragmented image; on the contrary,
large partitions result in a rough approximation of the edge.
We found that using either 12 or 24 partitions, depending on
the actual image, was appropriate in all cases we processed.
At the end of this stage, we have now a vector
representation of the edges, where each line segment is
International Archives of Photogrammetry and Remote Sensing. Vol. XXXI, Part B3. Vienna 1996
defined by its orientation, its gravity centre and its end
points.
Figure 3. Feature extraction output
Additional information on the goodness of the fit for
orientation and location is recorded; moreover, it is always
possible to go back to the original image region. Figure 3
shows the feature extraction output superimposed to the
original image.
SES Qe
Figure 4. The remaining features after data reduction
