Information and Knowledge Representation in Fusion Problems - Information Fusion in Signal and Image Processing - page 60

In image viewing and processing, for which we operate in environments that are
not completely specified and only partly known, we have examples of knowledge-
based systems essentially designed for supervising programs [CLE 93, NAZ 84,
THO 93] and for interpreting images [DES 90, GAR 89, HAN 78, MAT 86,
MCK 85]. Quite generally, they are comprised of the following components: isolated
points of the image, contours, areas. Semantic relations also come from artificial intel-
ligence: Part-Of , Is-a , Instance-of are the most common. It is much more complicated
to implement more complex structures, though they are quite useful to image pro-
cessing, such as concepts of hidden areas, shadows, multi-scale representation and
frequential components. Supervision strings together the expert's task based on the
objectives and the results that have already been obtained, and the distinction is made
between:
- systems guided by data, which put together, for example, pieces of contours to
form lines, then lines to form objects, etc., until they end up with a known object;
- and systems guided by goals, which start with the hypothesis of the result and
recursively search for the previous steps necessary to the presence of this hypothesis.

The prototype of such a system is well represented by Hanson and Rieseman's
VISION system [HAN 78, HAN 88]. A slightly similar method is found in many other
systems. The processing phase is separated in levels, as illustrated in Figure 5.2: typ-
ically a high level (level 3) which contains the symbolic descriptions and recognized
objects, then an intermediate level (level 2) which contains areas, lines, shapes, and
finally the low level (level 1) which contains images and their pixels. Operators use
the data on the level n 1 to create information on the level n . The other way round,

requests are made by the level n on the level n

−

1 in order to accomplish a certain
number of tasks necessary to decision making. Also, processes are constantly operat-
ing within each level to organize, arrange in order and fuse the information.

−

symbolic description of
the objects and scenes

inference,
deduction

level 3

focusing,
fusion,
hypothesis

matching,

perceptual
organization

organization,
grouping,
division

symbolic description
of areas, shapes
and lines

level 2

segmentation,
extraction

resegmentation,
redetection

stereo,
movement

level 1 image, pixels, data

Figure 5.2. Diagram explaining how a KBS works.
This representation is based on VISION [HAN 88]