Partial discriminative training for classification of overlapping classes in document analysis

Abstract

For character recognition in document analysis, some classes are closely overlapped but are not necessarily to be separated before contextual information is exploited. For classification of such overlapping classes, either discriminating between them or merging them into a metaclass does not satisfy. Merging the overlapping classes into a metaclass implies that within-metaclass substitution is considered as correct classification. For such classification problems, this paper proposes a partial discriminative training (PDT) scheme, in which, a training pattern of an overlapping class is used as a positive sample of its labeled class, and neither positive nor negative sample for its allied classes (those overlapping with the labeled class). In experiments of offline handwritten letter and online symbol recognition using various classifiers evaluated at metaclass level, the PDT scheme mostly outperforms ordinary discriminative training and merged metaclass classification.