Score Aggregation from Multiple Sources and Training in the Context of Lexicon Reduction using Holistic Features

Holistic methods developed for small, static lexicons are not easily extended to the large and dynamic lexicon scenario owing to word-level feature variability and paucity of training samples. A methodology of coarse holistic features and heuristic prediction of ideal features from ASCII is proposed to address these issues. The proposed methodology is based on the axiom that real-world examples of handwritten words may be viewed as the ideal exemplar of the word class distorted by the scriptor, stylus, medium and intervening electronic imaging processes. On a test set of 3,000 handwritten city names, we achieved a 70% reduction in the lexicon size (1,000) with 98.7% accuracy.