Fusion in Image Processing - Information Fusion in Signal and Image Processing - page 36

A radiologist in a hospital environment makes his decisions based on many pho-
tographs that often give different points of view of a same anatomical structure. These
photographs can be spread out over a view box and the doctor makes his opinion
based on closely and alternately examining several points of view. Based on a hypoth-
esis inferred from one of the photographs, he confirms his interpretations using the
others, clarifies it by cross-checking and including additional points of view, or on
the contrary rules it out based on contradictory information provided by one of these
points of view. This situation is typical of medical imaging, which is a field where
the acquisition techniques are growing more diverse: X-rays, magnetic resonance,
nuclear imaging and ultrasound imaging, each one leading to a variety of possible
modes depending on the acquisition protocols. It also finds support in the efforts made
by all hospital structures to group all of the image sources together in the same ward
or to have all of the images converge on a single console where the diagnosis will take
place. These efforts have progressively led to the introduction of integrated archiving
and consultation systems in hospitals (PACS 1 , for example).
For our second example, we choose a remote sensing expert whose task is to inter-
pret a complex scene. He has a large number of images at his disposal, provided by
various sensors, for example, images in the visible spectrum in different ranges of
wavelengths, or infrared images, or also radar images. Each source gives him infor-
mation on a particular aspect of the scene, thus allowing him to come up with a sce-
nario. Again, the expert works by confronting different representations, combining
them either to support his idea or to rule it out. His ability is the result of a consid-
erable amount of training and is increasingly complex as the image sources diversify
and grow in number. However, both satellite applications, for which many sensors
complete each other's information, and airborne applications, for which very different
sources are used (maps, cadasters, land occupation maps, geological or agricultural
surveys, elevation models, etc.), definitely tend to progress towards a greater com-
plexity of image sources.
In this context, image fusion appears as a task in itself, distinct from data fusion
because it is not clear whether it is possible to design an operational framework in
which every element of information would have its place, as was the case with the
satellite image that allowed us to include the “cloud cover” measurement in an overall
plan involving, for example, the evapotranspiration of vegetation cover and weather
conditions. In a broader context of image processing, image fusion is used to help
decision making in a complex and usually poorly formalized situation, in which the
different images provide an element of “truth” that contributes, in collaboration and in
opposition with other sources, to an overall interpretation. Therefore, by developing
automatic image fusion methods, preparing and shortening the human elaboration and
expertise phase, and possibly in situations with a large number of image sources, our
1. Picture Archiving and Communication Systems.