The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences. Vol. XXXVII. Part B4. Beijing 2008
333
calculation steps, the network generates binary images, with
value 1 pixels corresponding to the emergence of a pulse, and
value 0 signals corresponding to the lack of a pulse.
In research works described in Chapter 4, that network was
used as a generator of an aerial image fragment representation.
That representation is the so-called signature. It is a single
dimensional function, created after the application of the (4)
formula to subsequent binary images generated by the network:
G W=Z y «W (4 >
Uj
with n range of changes determining the length and location of
the signature.
3.2 Biological neural network model
Another type of image representation can be obtained after the
image has been processed with log-polar and log-Hough
transformations. Those transformations can be classified as
belonging to the broadly-understood class of neural networks,
as they are simplified models of visual information processing
that takes place on retina (Schwartz, 1977; Weiman, 1989).
The principle of operation of a log-polar transform is shown in
Fig. 2. A grid of concentric circles with exponentially
increasing radiuses is put on an image. Then, through a uniform
division of a round angle, the so-created rings are divided into
receptor fields. Information in receptor fields is averaged, so the
log-polar representation is the most precise when close to the
image centre (near the so-called dead zone, that is a non-
transformed area), and less precise when close to receptor fields
located far from the centre. Main advantages of the log-polar
transformation include a considerable reduction of data, as well
as transformation of image rotation and magnification in the
Cartesian space to parallel shifts along the respective axes in the
log-polar space.
Using the log-Hough transform, it is possible to detect sections
of straight lines and gentle arcs on log-polar edge images
(Weiman, 1989). In the log-Hough space, there appear maxima
in places that correspond to the parameters of detected lines.
Those maxima are the higher, the more collinear points were
found in the image in the log-polar space after the separation of
edges.
receptor
field
blind spot
fragments of lines with different inclinations. Both those
transforms were selected to design the image representation,
because they reduce information and provide condensed
information on the location of edges in relation one to another.
4. SCOPE OF RESEARCH
The conception of a large-scale research on the use of neural
networks in photogrammetry is shown in Fig. 3. Bearing in
mind the division of image processing techniques into those
"object-centred" and "area-centred" ones, the authors focused
on the latter ones.
sub-image
Figure 3. Flow chart of the experiments. Neural-like methods
are denoted in italics
The left fragment of Fig. 3 includes two methods that are based
on the utilization of gradient images. When analysing the right
part of the drawing, one should note that the “raw” image
constitutes an input both to log-polar transform and ICM
network.
Fig. 3 also shows, at the bottom part, the goals set by the
research team. The first of them included the creation of aerial
photo sub-image representations, which were later used for
selection (the second goal) and, at the same time, for matching
(the third goal). It is that second goal that requires a separate
discussion here.
Figure 2. Distribution of receptor fields in the log-polar
transformation
In the course of a photogrammetric analysis, one indicates
informative points on one of the images, seeks out first a larger
area on the second image, and then equivalents of points
indicated on the first image. The second goal (selection)
The operation performed by the log-Hough transform is similar
to the transformation observed in the visual cortex: detecting