|
Abstract:
|
Multi-class classification is the core issue of many pattern recognition tasks. Several applications
require high-end machine learning solutions to provide satisfying results in operational contexts. However,
most efficient ones, like SVM or Boosting, are generally mono-class, which introduces the problem of
translating a global multi-class problem is several binary problems, while still being able to provide at the
end an answer to the original multi-class issue.
Present work aims at providing a solution to this multi-class problematic, by introducing a complete
framework with a strong probabilistic and structured basis. It includes the study of error correcting output
codes correlated with the definition of an optimal subdivision of the multi-class issue in several binary
problems, in a complete automatic way. Machine learning algorithms are studied and benchmarked to
facilitate and justify the final selection. Coupling of automatically calibrated classifiers output is obtained by
applying iterative constrained regularisations, and a logical temporal fusion is applied on temporal-redundant
data (like tracked vehicles) to enhance performances. Finally, ranking scores are computed to optimize
precision and recall is ranking-based systems.
Each step of the previously described system has been analysed from a theoretical an empirical
point of view and new contributions are introduced, so as to obtain a complete mathematically coherent
framework which is both generic and easy-to-use, as the learning procedure is almost completely automatic.
On top of that, quantitative evaluations on two completely different datasets have assessed both the
exactitude of previous assertions and the improvements that were achieved compared to previous methods. |
