Image Processing Reference
The complexity is also a result of the strong heterogenity of information to com-
bine, whether it is images taken from different sensors, images and models, or different
characteristics extracted from one or several images. This disparity is found both in
the nature of the information and in their representation. The data can be frequent,
for example, when dealing with a typical case for a given application, for which it
is possible, for example, to obtain statistical data. It can also be scarce (for exam-
ple, pathological images) and in this case it is much more difficult to model them in
a statistical way. The combination of these two types of data is common in image
processing. Furthermore, they can be factual (typically, a photography of a scene at
a given time) or generic (a model, rules, general knowledge about the application).
Combining elements of information with different specificities often leads to conflict
problems to solve. In image processing, this is not an easy task because factual infor-
mation is not always reliable and accurate enough for it to systematically be given the
priority over less specific and more generic information that may allow exceptions.
The combination of information is often guided or constrained by additional infor-
mation regarding the information to combine, the context and the field of application.
It is also a source of strong heterogenity. One example of additional information on
the information is the reliability of a source, either overall or conditional to the objects
observed. This is a very common case in the classification of multi-source images,
where an image may be reliable for one class but not for another. Here are a few
examples of additional information about the subject and the context:
- rivers are dark in SPOT's XS3 channel (information relating the type of acquisi-
tion with an observation);
- the CSF is dark in MRI images in T1;
- roads cross each other to form intersections (integrity constraint).
This generic information is used to guide the fusion process. The last example is a
typical case of a rule with exceptions. The rule gives the most general case, but is not
true in the case of dead-ends, for example.
Active fusion is one of the means for reducing complexity by choosing at every
instant the best information to fuse. This choice can be performed based on a partial
result of fusion obtained at a previous stage, on information measurements, on outside
information likely to guide the fusion, on the identification of ambiguities that have to
be cleared up, etc.
In image processing, fused information is necessarily tainted with imperfections
(uncertainty, imprecision, incompleteness, ambiguity, conflict, etc., according to the
distinctions proposed in Chapter 1). These imperfections originate on different levels,
from the observed phenomena to the processes. For example, the smooth transition
from healthy tissue to pathological tissue is an imprecision caused by the physio-
logical phenomenon. Likewise, similar characteristics between two different kinds of