Feature identification for hybrid structural/statistical pattern classification

TLDR: Large-scale statistically-significant trials, in the context of a mixed-font, variable-size optical character recognition (OCR) system, have shown that the technique is superior to simpler, fixed mappings, and is effective in generalizing common characteristics in mixtures of fonts.

Abstract: A general technique for combining the strengths of structural shape analysis with statistical classification is proposed. The approach is to construct a function, called a feature identification mapping, from the representation generated by structural analysis to the one required for statistical classification. It is shown that if a certain continuity property holds for the parameterizations of the structural shape types, then it is possible to infer the mapping automatically. Inference is slow and heuristic, but is highly automated, controlled by only a few statistical parameters, and is applicable uniformly to all shape types. In addition, if the shape types are sufficiently elementary, the resulting mapping can be computed quickly using kD-trees. Large-scale statistically-significant trials, in the context of a mixed-font, variable-size optical character recognition (OCR) system, have shown that the technique is superior to simpler, fixed mappings, and is effective in generalizing common characteristics in mixtures of fonts.