In on-line English word recognition, the performance depends much on the size of lexicon, i.e. the number of possible choices. In this thesis, we propose a method of lexicon filtering by using a set of robust global features of on-line English word.
The features include dot, horizontal bar, number of down-strokes, and type of down-stroke. The number of down-strokes is used as a rough estimate of word length. The sequence of down-strokes, classified by artificial neural network, are effectively used as the overall shape of the word. In case of using all these features combined, we could reduce a given lexicon down to 1% with 3% error on the average. And the performance in on-line English word recognition was improved by using the proposed lexicon filtering method.
