An engine for cursive handwriting interpretation

Abstract

This paper describes an engine for on-line cursive handwriting that requires very little initial training and that rapidly learns, and thus adapts to, the handwriting style of a user. Key features are a shape analysis algorithm that efficiently determines shapes in the handwritten word, a linear segmentation algorithm that optimally matches characters identified in the handwritten word to characters of candidate words, and a learning algorithm that adds, adjusts, or replaces character templates to adapt to the user writing style. In tests, the system was trained on four samples of each character of the alphabet. One writer wrote these samples in isolation. Using a lexicon with 10,000 words, the system achieved for four additional writers an average recognition rate of 81.3% for top choice and 91.7% for the top three choices. The average response time of the system was 1.2 seconds per handwritten word on a Sun SPARC 10 (42 mips).