Image Processing Reference
10.1.2. Design constraints and concepts
The vision problem poses a number of difficulties:
- we do not know beforehand which processing chain can solve the vision problem
we are dealing with. It is chosen according to the context and the situations encoun-
tered. There is never an overall objective defined, on the contrary, local objectives
come up depending on the processes being conducted. Furthermore, the aspect vari-
ability for a given object or several objects in an image suggests that it is impossible to
apply a comprehensive process to the entire image. Thus, the system, depending on the
context, needs to adapt its strategy, the operators as well as the parameters according
to the local information that is gathered;
- the specification of entities, objects, classes and sub-classes needs to allow a high
variability within classes and from one class to another. Of course, this implies that
the systems need to be capable of handling uncertainty;
- the information we are interested in is sometimes concentrated in a few pixels,
for example, in detection and it is drowned in a considerable amount of data, hence
the need to follow an incremental, safe and progressive procedure in order to gather
knowledge. The objective is to rely on results that have already been obtained to con-
tinue with the process;
- the context of a multi-target application imposes the need for a monitoring alert
function that can very quickly inform the system of the presence of a possible threat.
All of these difficulties described here raise a number of questions regarding design
that have been discussed in other works, such as the ability to adapt, to focus, to
distribute, to handle uncertainty and the system's incrementality.
10.1.3. State of the art
Over the past two decades, there has been considerable progress on the subject
of DRI from a scientific and technical perspective. Bhanu and Ratches have drawn
a particularly detailed review of this progress [BHA 86, RAT 97]. For the most part,
the systems suggested in the early 1980s were heuristic. They would typically use
the following sequential scheme: segmentation, detection, extraction of parameters,
classification and tracking targets.
At the end of the 1980s, a new generation of systems appeared, with the inten-
tion of breaking this sequentiality. These methods explicitly integrated the knowl-
edge and techniques of shape recognition (knowledge-based vision and model-based
vision). They usually go through the following phases: searching for relevant areas,
then recognition and identification of targets in a more limited area [BHA 92].
Today, improving performances requires associating complementary information
in order to provide an adequate response to the operational needs of situation analysis.