Real time recognition of hand-drawn characters

Abstract

This paper describes a system designed to recognize handdrawn characters in real time. The central feature of the system is its use of the time sequence information of the input character. It its abstract form, the system may be viewed as a collection of discrimination nets, or filters. Each net operates on the input character, or rather on a sequence of property vectors representing the encoding of the input character in the time domain, and produces a set of so-called candidate characters. The system then utilizes reliability estimates for the individual nets to select the final output character. A program representing a particular implementation of the system has been written for the DEC PDP-1. The user draws the character on the face of a cathode ray tube with a light pen. The program follows the pen and constructs the appropriate sequence of property vectors. The properties used are simple geometrical ones, and the descrimination nets are tree structures which store sequences of property values. Recognition time is of the order of .25 seconds, of which approximately .15 seconds are occupied by drum read-and-write operations required by the small memory size (4K) of the machine. The user can teach the program to recognize his set of characters. The learning process for the program involves modifying individual decision trees, changing the weights on each tree, and where necessary, introducing new decision trees with their corresponding properties into the system. Because the program was written as an input device to a larger man-machine system, the description of the implementation stresses the human engineering features. A qualitative evaluation of the system as implemented is offered, together with possibilities for expanding and generalizing the program.