Multiobjective selection of features for pattern recognition

The paper suggests a novel pattern recognition system based on a flexible genetic selection of relevant features. Firstly, a hybrid set of competing features is determined, aggregating the results provided by several different basic extractors, such as principal component analysis, bi-dimensional Fourier transformation, grey-levels and geometric analysis. Subsequently, the most suitable features are chosen, in accordance with the specific properties of the particular visual patterns that have to be recognized, via a multiobjective optimization performed in terms of classification accuracy, parsimony and computational requirements. Pareto-optimal solutions are searched using genetic techniques based on hierarchical encoding. To adapt the selection pressure imposed by the conflicting objectives, a new algorithm for fitness computation is proposed. It efficiently exploits the concept of dominance analysis due to a progressive articulation between the decision mechanism and the search procedure. The experimental trials, performed within the context of a holonic palletizing manufacturing system, illustrate enhanced adaptation capabilities of the designed pattern recognition subsystem.