The Character Recogniser

The underlying character recogniser used by this system is the one used by Smithies, Novins, and Arvo . Symbols are encoded as collections of polylines representing individual user-drawn strokes. The recogniser uses an extremely fast on-line recognition algorithm based on nearest-neighbour classification in a feature space of approximately 50 dimensions. Rubine  and Avitzur  both use a similar feature-based strategy.

To train the character recogniser, the user supplies ten to twenty handwritten samples of each character. These samples are stored and used by the recogniser to recognise the input characters.

Although the recogniser is theoretically user-dependent, the system is relatively user-independent in practice. For example, even though the recogniser was trained using samples supplied by just two people, others had little difficulty in using the system.

For each group of strokes passed to the character recogniser, the top n interpretations of those strokes are returned, along with a confidence for each interpretation. This confidence information is important as it is used by the stroke grouping method, described in Section .

To get higher recognition rates from the character recogniser, and thus improve the performance of the stroke grouping, more versatile classifiers, such as neural nets , and perhaps the use of contextual information from later processing stages, as described by Miller and Viola  could be used.

Virtually any character recognition module can be incorporated into this system. The only requirement imposed by the system on the recognition module is that it must be capable of ranking the n most likely candidates for a single pattern by a numerical measure of confidence, and that the confidence measures of different patterns must be directly comparable.

could be used.

Virtually any character recognition module can be incorporated into this system. The only requirement imposed by the system on the recognition module is that it must be capable of ranking the n most likely candidates for a single pattern by a numerical measure of confidence, and that the confidence measures of different patterns must be directly comparable.