Handwritten character recognition is a highly challenging area in the field of pattern recognition. In performing recognition, any single classifier system has its strengths and weaknesses. The objective of this dissertation is to develop multiclassifier systems which utilize the combined strength of several classifiers to make a significant improvement in recognition over the single classifiers. The multiclassifier systems developed were cascaded, vote-to-decide, confidence enhancement, and hierarchical learning systems. In each multiclassifier system, the single classifiers contained their own feature extraction, similarity measure, learning, and classification stages.
Feature extraction extracted object features and formed feature representations. Three feature representations were developed, which were the angle sequence, vector contour representation (VCR), and Fourier transform representation (FTR). To evaluate the similarity of objects in different representation forms, measures based on Euclidean distance, vector correlation, string matching cost, and Fourier transform were developed.
For learning, two supervised clustering techniques were developed: maximum region clustering (MRC) and accumulated potential clustering (APC). The MRC learning maximized the clustering regions by including as many samples of the same type as possible in each cluster without enclosing any alien sample. In APC learning, the feature space was viewed as an electrostatic field in which each cluster served as a potential generating center. Each object class established the minimum number of cluster centers necessary to protect its members from being attracted to other classes.
In the classification stage, the MRC classifiers identified a test sample with the class of its nearest cluster center. The APC classifiers assigned a test sample to the object class which attracted it the most. In a multiclassifier system, the final classification decision was made based on the individual decisions made by the single classifiers. The rules of making a final classification decision depended on the multiclassifier system and differed from system to system.
All the developed single classifier and multiclassifier systems have been simulated on a Sun computer, and the results were presented. The computer simulation results showed that the multiclassifier system significantly improved the recognition performance.