How Computers Recognize - Robots Unlimited: Life in a Virtual Age - page 118

Figure 29. The handwritten word “Largo” and its representation in Parascript elements
(Courtesy of Parascript R ,LLC)
of eight elements that are sufficient to describe all the trajectories of the
pen found in the cursive letters of the Roman alphabet (see Figure 29 ) .
During the recognition process, NHR employs a reference vocabu-
lary—a type of dictionary—to provide information that relates to the
context of the writing. For example, the system might be trying to recog-
nize whether the image of a particular word is “clear” or “dear”. If one
of these images occurred in a form where the writer was asked to spec-
ify the colour of the lenses in his spectacles, the software would look up
both words in its reference vocabulary and eliminate the word “dear” as
being an inappropriate response, and select “clear”, which is the more
appropriate word in that context.
In order to increase the accuracy of its word recognition, NHR em-
ploys two types of word recognizers, called handwritten and analytical .
The handwritten word recognizer deals with the word as a single, unseg-
mented unit, in which the word is represented by a series of elements.
This series of elements is matched to a corresponding word in a dic-
tionary and, as part of the matching process, the system provides a con-
fidence level to indicate how certain it is that it has chosen the correct
word from the dictionary.
The analytical word recogniser is used mainly for numerals and print-
ed characters. It employs two classification methods—one is based on a
database of symbol prototypes and the other comprises a number of arti-
ficial neural networks. 5 The combined recognition results of both types
of classifier, the database classifier and the neural network, indicate the
character recognized by the software and provide a confidence measure
for that character.
Writer Independence and System Adaptability
Handwriting recognition systems depend on a stored database of charac-
ters and words that provide the “training” for the system. Some systems,
called writer-independent systems, are trained on writing collected from
5 See the section “Artificial Neural Networks” in Chapter 6.