Classifier design with incomplete knowledge

摘要

Pattern classification is a well studied problem in which the identity of an unknown pattern is determined to be one of M classes spanning the pattern space. A new criterion function, the Inck (Incomplete knowledge) criterion function, is proposed which approximates the error probability when the M classes do not span the entire pattern space. This criterion is based on the probabilistic measures obtained from a modified version of Dubuisson and Masson's statistical decision rule with reject. The error probability, or conversely the probability of correct classification, can be determined without having complete knowledge about the class distributions. The attractiveness of this criterion is that it is highly correlated with the statistical decision rule (with reject) which can be used by the classifier. Thus, the criterion provides a good indication of the classifier performance that can be expected with the reduced feature space.A classifier design is presented which incorporates the new criterion function. As well, it is described how the rejection classes of the statistical decision rule with reject can be used in a multiresolution classifier design to determine the class identity of patterns which cannot be determined at a given resolution. Experimental evidence is provided which demonstrates the advantages of the Inck criterion function used in conjunction with a multiresolution classifier design.