Improved Character Recognition through Subclassing and Runoff Elections

A succinct representation of handwritten symbol curves can be obtained by computing truncated Legendre-Sobolev expansions of the coordinate functions. With this representation, symbol classes are well linearly separable, which yields fast and robust classification algorithms based on linear support vector machines. We have found that the presence of different variants of a symbol in the same class reduces linear separability and correct retrieval rates of the linear SVM classifiers. We show experimentally that by labeling the variants we obtain nearly 100% linearly separable classes and improve the correct retrieval rates by about 3%. This increases the number of classes, however, and the consequent effect of irrelevant classifiers. We introduce a general technique to correct this effect by replacing the conventional majority voting scheme with a runoff election scheme. We have found that such runoff elections further cut the top-1 mis-classification rate by about half.