Fusion in Image Processing - Information Fusion in Signal and Image Processing - page 43

spatial context, or on a more structural level by way of spatial relations between the
structures or the objects of the scene.
The fact that the data is growing in volume can exert some constraints on the
calculation time. Thus, handling single pixels is limited to simple operations. More
complicated operations often require including information of a higher level and a
more structural representation of the information.
The complexity and amount of data dealt with often requires a choice of the infor-
mation and knowledge to fuse. This choice is of course guided mainly by relevance
criteria with regards to the decision objectives, but also by criteria involving the diffi-
culty in accessing and representing information and knowledge, as well as their qual-
ity.
Finally, evaluation is a crucial problem. In image fusion problems, there is usually
a “truth” or a “right” solution, but it is often difficult to find. Thus, evaluating and val-
idating a fusion method can only be done by using simulations, phantom acquisitions,
or by comparing with a manual decision. This situation is different from problems of
vote or social choice, where there is no truth and the goal is to find a “best” solution
expressed as a compromised, based on equity and ethics criteria.
3.5. Numerical and symbolic aspects in image fusion
If the numerical aspect of information handled in image processing is obvious, its
symbolic aspect deserves a little more attention.
The symbolic information can be bound to the data we wish to fuse (for example,
visual information in a map or an anatomical atlas, attributes calculated using the data
or objects previously extracted from the data) or to the knowledge of the field. Typi-
cally, the information on the field is often represented by rules, structural representa-
tions such as graphs, often used in shape recognition in images, constraints that need
to be taken into account in the algorithms. Structural information can, for example,
specify that a road network can be represented by a graph using roads and intersec-
tions, or it can express in the form of propositions general rules about the scene such
as “brain ventricles are always inside the white matter”, etc. Structural information
can also be represented in the form of icons and therefore in a way similar to images,
for example, in the case of digital maps or anatomical atlases. The geometric represen-
tations of the structures are then combined with the nature or the semantics of these
structures. However, the information in this field can also be purely numerical when it
consists, for example, of specifics about the acquisition, such as the wavelengths used
in satellite imagery, or acquisition times in medical imaging. Hybrid representations,
in which numbers are used as symbols with a quantification, are used in image fusion
to quantify the quality of a sensor, to evaluate symbolic data or the confidence in a