Image Processing Reference
Among other things, they make it possible to improve the robustness of the analysis
with respect to its environment, the acuteness of the information transmitted and the
reactivity in time. However, a smart use of the synergies between the different sources
of information and the volume of available data requires the implementation of data
fusion methods. Different theoretical methods would seem to satisfy this condition
since they can be used, for example, to deal with uncertainty, imprecision, incomplete-
ness, reliability, dependence or relevance. Among these methods, we should mention
mainly probabilistic [BUE 97], possibilistic, fuzzy set and evidence theories, as well
as logical or connectionist methods [HEN 93, ROT 90, YOU 98].
The integration of multi-sensor data, of spatial and temporal knowledge has
opened new prospects for the future. The goal is no longer to develop new algorithms
but rather to take advantage of partial and imprecise information provided by each
one of them. The most common approach remains the prediction and verification of
hypotheses [BEV 97, WU 97]. The major drawback is often the handling of a great
number of hypotheses, which imposes a large number of combinations to examine.
This method of exploration, associated with feedback, has the advantage of question-
ing hypotheses, and of choosing algorithms and parameters [DRA 95]. In the event of
poor performances, it authorizes a change of strategy rather than pursuing the analysis
with faulty hypotheses. It allows the system to adapt itself to the current situation.
The calculation of a cost/benefit function makes it possible to complete this adaptive
scheme by asking the question of how useful an action is, knowing the current situa-
tion and uncertainties [ROB 94].
Constraints tied to the heterogenity and the amount of knowledge naturally lead
to architectures that distribute knowledge. The objective from now on is to equip the
DRI system with the ability to manage, select and activate its own resources so as to
conceive efficient control strategies to go through with the various analysis phases.
Today, a DRI system is at the intersection between image analysis and cognitive rea-
10.2. Proposed method: towards a vision system
The problem of vision is stated as an incremental problem of information gath-
ering, in which at each step of the process we ask the questions “where, what, how”
[GAR 00, RAO 95]:
- where is the relevant information located in the image?
- what is the relevant information?
- how is this information extracted from the image?
In order to answer these questions and satisfy the constraints stated before, a vision
system must constantly rely on the information gathered to construct its own knowl-
edge of the field. The objective is to suggest new areas of focus, different goals and