MELIDIS: Pattern recognition by intrinsic/discriminant dual modeling based on a hierarchical organization of fuzzy inference systems
Résumé
In this article, we present a new recognition approach which
aim is to combine properties that are rarely fully satisfied in the
same classifier: performances, robustness, genericity,
compactness and transparency for the designer. This last point makes
the system easier to maintain and optimize
for a given application. The classifier is totally data
driven and its architecture is based on the properties of the knowledge used. The main
originality comes from a specific cooperation of intrinsic and discriminant
knowledge. The system is organized in two levels: the first one models
classes with intrinsic fuzzy prototypes and the second one operates
a discrimination by fuzzy decision trees. To improve the
discrimination process, this one focuses on contexts
induced by a focus mechanism based on intrinsic knowledge.
For transparency, the system is formalized by fuzzy
inference systems combined for decision. Experiments
on severals problems have shown that the system's performances are
close to SVM ones, with 10 to
30 times less parameters. They have also demonstrated the interest of the
collaboration between intrinsic and discriminant knowledge.