Japanese handwriting recognition using AI techniques

Abstract

An online computer system has been developed to recognize handwritten Japanese Hiragana characters. The system requires only detection of pen-up and the velocity of the pen in the two horizontal planes of the paper. No absolute coordinates need be collected. This enables a user to write naturally on paper while recognition takes place, using a biro pen modified with a rubber membrane and three strain gauges. Such tolerance towards the user creates scope for considerable ambiguity, techniques from artificial intelligence, in particular from natural-language processing are used to help with this problem. A grammar is used to describe the target character shapes, and the input stream is classified by parsing. For characters close to the target, the most likely of the hypotheses was correct for over 95% of the characters drawn by three native Japanese. For poorly drawn characters the recognition rate, based upon the most likely hypothesis, drops to around 80%, but the set of hypotheses almost always contains the target character.<>