Simplifying the model-based classifiers for multi-modal problems in classification of spectra

Industrial applications of spectral imaging introduce novel types of problems where generally applied assumption of class unimodality does not hold anymore. Object sorting problems, for example, become multi-modal as soon as the terminal highlevel classes are defined as ad-hoc collections of material types. This paper discusses design of model-based multi-modal classifiers in problems, where the class modes are apriori known during training. We consider derivation of data representation and classifier design as connected issues. Several designs simplifying the full Gaussian mixture-based model using diverse representation building strategies and regularization are discussed. Apart of model-based algorithms employing mode descriptors such as Gaussian mixtures or multi-modal SIMCA, also an alternative method based on inter-mode discriminants is considered. A set of experiments is conducted on an artificial dataset modeling some aspects of multi-modal spectral classification problems and two real-world datasets from industrial objectsorting application. The behaviour of different model-based methods is studied using learning curves. The main conclusion of this paper is that incorporation of supervised information significantly improves classification performance, reduces the complexity of the final system and speeds-up its execution.