EFFECTS OF NOISE IN CHARACTER RECOGNITION SYSTEMS.

Abstract : In this report, a technique for evaluating character-recognition systems is developed. First, a general mathematical model for pattern-recognition systems is specified. Then it is shown that various aspects of system and subsystem performance can be evaluated by considering the systems and subsystems as simple communication channels. The evaluation results in a statistic, termed the 'intelligence coefficient,' which is a measure of the fractional amount of source-generated intelligence available at the output of a communication channel. It is shown that the intelligence coefficient can be used to do the following: to predict the limitations imposed on system performance by input-signal conditions (e.g., the limitations on character recognition of carbon copies); to measure the effectiveness of signal transformations; and to measure the effects of decision logics. As one means of acquiring the signal data necessary for calculating the intelligence coefficient, empirical noise curves and analytic models for the distribution of ink noise about a character are presented. It is recommended that the intelligence coefficient be programmed for computation on a general-purpose digital computer and be used as an analytic tool for the evaluation of proposed recognition systems and for the design of improved recognition systems. (Author)