Open Collections

A pen-tip pressure-based system for real-time handprinted character recognition

University of British Columbia

An on-line handprinted character recognition scheme was developed, implemented, and tested. The selected symbol class included the numerics subset of the ANSI standard and nine special symbols defined for a proposed dedicated graphics editor system. A primary design objective was to provide an environment that would reflect the non-technical user's natural tendencies. A Xebec ball-point pen with a pressure-sensitive tip monitored drag forces generated in the plane of the paper. Force information was available to the processor by periodically sampling a strategically placed strain-gauge array. Data were analyzed in real-time to provide immediate inferences concerning stroke motion. Since positional information was not available no constraints were placed on the writing medium. The system is microprocessor compatible with total software comprising less than 1500 16-bit words. It is also very CPU efficient owing to a multi-level pre-processing scheme that significantly reduces the computing burden. Feature extraction is simple and effective. A hysteresis scheme is applied at all borders of the quantization plane to help eliminate redundant data. In the final stage input symbols are classified with the aid of a modified tree structure that is inherently tolerant to a large variety of local variations. The dictionary is compact and quickly traversed; provision for 19 symbols requires less than 250 words since only a fraction of all allowable forms need be explicitly specified. Extensive testing was conducted on the completed recognizer. A small amount of user training is required to ensure adequate performance. A total of more than 4000 characters obtained from 15 subjects were analyzed. Experienced users achieved a 96.3% average recognition rate and scores above 98% were not uncommon. Performance is independent of character size and position and users appear to retain skills acquired during training. Nontechnical and technical users perform equally well and are not distracted by the small amount of hardware involved. Efforts were directed to determine if the system could be dedicated to a particular user. In some cases, it was found that dictionary reconfiguration improved performance markedly. Optimal system parameter values were fairly robust, optimizing for a single subject yielded good results for all subjects. Exceptional writers, however, might benefit from some parameter fine-tuning. An automated personalization prototype designed to select optimal values was developed and tested.

Text

2010-03-06

For non-commercial purposes only, such as research, private study and education. Additional conditions apply, see Terms of Use https://open.library.ubc.ca/terms_of_use.

http://hdl.handle.net/2429/21603

Electrical and Computer Engineering

University of British Columbia

Graduate