TROLOGY
nnes@avt.at
au,
rs. The emphasis
ms and processes
of the automatic
aspects covered
port initial image
gration of these
ntal models and
aspects of these
determination, a
tion, as well as a
> tests. A more
2001).
ated VM process
] targets in each
1s a significant
elligence, the use
simplified the
n algorithms for
image scanning
e quality. Under
argets will yield
, as indicated in
ontrast difference
be as distinct, for
targets or when
essary to account
Figure 1b, as far
age mensuration
dation process is
their nature can
tric triangulation
mage recognition
jid measurement
th the discussion
only. The strategy
d target quality
ncluding distinct
e lc shows that
n the targets and
their background, all have a common feature, namely distinct
edges. Figure 2 shows the respective intensity profiles through
these targets. Although the distinct jump in grey values at the
edge of the targets varies in magnitude, the edge gradient in
each case is more or less the same (recall that we design targets
to yield such edge gradients). Thus, the gradient is generally a
better criterion for the detection of targets than the commonly
adopted simple pixel-value thresholding (e.g. Atkinson, 1996;
Luhmann, 2000).
Figure 1. Target images in VM: high-quality, near-binary (a),
low-quality (b) and a range of images (c).
The grey value gradient is expressed via the first derivative of
the intensities. With the image being interpreted as a two-
dimensional discrete function, the first derivatives in row (x)
and column (y) direction can be given as
"x t
8;,;78,j178i;j
"P i
fi Sins Si
(1)
In the literature, alternative operators are also described (e.g.
Gonzalez & Woods, 1992), with some using more than two
neighbouring pixels to compute the gradient at a certain
position. However, derivatives using only two adjacent pixel
values are well suited to near-circular targets, and they have the
advantage of simple and fast computation. Performance
30 30
20 az
20 20
Lu m
40 10
D m
0 0
considerations are very important for VM scanning algorithms,
with higher contrast imagery generally leading to optimal
computational speed.
2.2 Image Segmentation
Given the edge characteristics discussed, a line-orientated image
segmentation algorithm that uses the gradient as a target
recognition criterion would appear as a good candidate for VM
applications. The gradient criterion alone, however, can deliver
unsatisfactory image segmentation results in conjunction with
low-quality images, since brighter areas are always recognised
as targets if they are surrounded by darker regions. To alleviate
this problem an extended segmentation algorithm has been
developed for Australis, which uses both the gradient criterion
and a minimum grey value threshold to detect target regions.
In comparison to algorithms that only use minimum grey value
thresholding, the value of the adopted minimum threshold in the
approach described is quite low. The threshold is not used so
much to differentiate between isolated bright areas and targets.
Instead, it only guarantees that pixels below a certain grey
value, which are definitely non-target pixels, are rejected from
further target measurement processes. The two thresholds used
in the scanning process, the gradient threshold and the grey
value threshold, need to be carefully chosen to ensure
satisfactory results in the segmentation process. A fixed
gradient threshold is generally appropriate for quasi-binary
images as well as for low-quality images. However, the grey
value threshold has to be separately selected for each image
using histogram analysis.
With this development, the line-orientated scanning algorithm
of Australis is implemented as follows: The process starts at the
first pixel line of the image and goes through every pixel in that
line. If the current pixel position fulfills both the grey value and
gradient threshold criteria, the beginning of a target blob is
indicated. The end of the target blob is detected either if the
pixel grey value is below the minimum threshold or the gradient
threshold is exceeded in the negative direction. If the ‘negative’
gradient criterion is detected at multiple positions, the last
position is taken as the end of the target. It may happen that the
criterion is met, without the detection of a target beginning. In
this case, if the actual pixel is above the minimum threshold, the
start of the target region has obviously been missed. To find the
beginning of the missed blob, the algorithm searches backwards
until the pixel grey value drops below the minimum threshold.
Figure 2. Intensity profiles through the targets of Figure 1c.
-61-