Evaluation of recognition accuracy for discrete and run-on writing

Abstract

In an effort to reach markets in which a keyboard/mouse interface is difficult or inconvenient to use, manufacturers are now beginning to introduce light-weight portable computers which recognize hand-printed characters. The recognition accuracy that these new computers are able to achieve will be a critical factor in determining their acceptance by users. There are, however, few published studies of handwriting recognition accuracy and the variables which affect accuracy. The purpose of this study was to assess recognition accuracy as a function of a number of factors which might vary in the real-world use of handwriting recognition systems. These factors included style of writing, amount of training, interval of disuse, and alphabet. The findings suggested that recognition accuracy reached a steady-state level with a relatively small amount of training and remained at that level for as long as a month (the longest interval tested in this study). For an 82-character alphabet, character recognition accuracy was 92.7% for discrete writing and 87.6% for run-on writing. Accuracy with alphabets restricted to digits or uppercase only was quite high, with mean recognition rates ranging from 95% to 99%. The implications of these findings for the design of real-world recognition systems are discussed.