Image Processing Reference
- involving the quality of the information: two images of the same type, but dif-
ferent acquisition parameters can lead to elements of information of different qualities
for different structures.
Redundancy is caused more obviously by the fact that the images we wish to com-
bine describe the same scene. For certain fusion problems, such as group studies in
functional imaging, redundancy (which areas are activated by all of the subjects) and
complementarity (where are the differences) are subjects of study in themselves.
Conflict is a very delicate matter, as with other applications of information fusion,
as we discussed in Chapter 1. With images, examples of conflicts that are only appar-
ent and easily confused with complementarity occur when an image is not capable of
distinguishing two classes whereas another one can. Imprecision and uncertainty are
also sources of conflict. For example, a poorly localized contour can cause a conflict
between several contour detectors. Conflicts due to the different specificities of the
elements of information to combine are common in image and model fusion applica-
tions. For example, recognition of brain structures by fusion of MRI images and data
found in an anatomical atlas must deal with variability among individuals, which is
often not represented in the atlas, or also the possible presence of pathologies in the
patient's images, which are not found in the generic model. Similar problems occur in
the fusion problems of aerial and satellite images with digital maps. In this case, the
conflict can be due to an imprecise drawing of the map, to modifications of the scene
not included on an older map, etc.
There are several types of constraints specific to image processing that have to be
taken into account.
From the perspective of the fusion system's architecture, decentralized systems
are rarely imposed. The most common case is that of off-line fusion, in which all the
elements of information are available simultaneously. Centralized systems can then be
Real-time constraints are fairly rare, except with surveillance or multimedia appli-
cations, in which they are destined to play a growing role. We will come across such
constraints again, but in a much stronger form, in the parts of this topic that describe
fusion in robotics, for example.
On the other hand, spatial consistency constraints are very stringent and constitute
an important subject of research in image fusion. An increasing amount of studies
focus on taking into account spatial information, either on a local level by way of the