(A) multiple classifier combination scheme for highly reliable handwritten numeral classification

Abstract

For the purpose of commercialization of Handwritten Numeral Recognition Systems(HNRS), it is desirable that the HNRS produces less misclassification rather than more correct classification by introducing a rejection scheme. In this dissertation, we propose the Hierarchical BKS as one of the multiple classifier combination frameworks. It utilizes ranked level individual classifiers, and, automatically expands its behavioral knowledge in order to satisfy given reliability requirement. To enhance generalization capability, we introduce the concepts of UOT and marginalizing of BKS. The meaning of UOT is how much we believe that the unit occurred in the training stage is present in real situation. Marginalizing is a utilization of BKS constructed by every combinations of individual classifiers in the pool, except all classifiers. Several comparisons of Hierarchical BKS with BKS and unanimous voting are shown and characteristics of Hierarchical BKS are discussed. As the results, Hierarchical BKS outperforms BKS with both modelling and unseen data, and, only Hierarchical BKS delivers reasonable classification ability as a practical MCC frame-work. Hierarchical BKS has an excellent modelling ability : When BCRT is set to 0.9995, Hierarchical BKS is able to classify all modelling patterns. Also, it is con-firmed that the BCRT is a lower bound of the reliability in the modelling data. We can see that the problem caused by different distribution of modelling and unseen data is overcome by increasing the value of UOT, and by the classification algorithm using marginalizing of BKS. By doing that way, the difference between the assessments of performances with modelling and unseen patterns becomes smaller, which means UOT is a very useful threshold to overcome problems caused by the difference of the samples and population.