Key points are not available for this paper at this time.
Design of pattern recognition systems usually involves a number of uncertainties which can be resolved only by experiment. In statistical recognition systems, pattern information is stored in the machine in the form of pattern characteristics with statistics relating the characteristics to the patterns. If the number of characteristics a system can process is limited, and if the system designer, while able to conceive of a large number of relevant characteristics, does not know which are the most important, then an experimental selection of these characteristics is required. As tools, the designer may have at his disposal a large computer, a large sample of the patterns to be recognized, and a set of programs to measure the characteristics. Thus, he can compare certain statistics relating the patterns to the characteristics. The problem is: what statistics should he calculate in order to select the best characteristics? Assuming the characteristics to be independent in their effect on the decision, this paper examines the notion of a single number statistic for each characteristic which would have certain desirable properties related to the "goodness" of the characteristic. It is shown that, in general, no such statistic exists. However, a statistic is proposed which, while not having these properties in general, at least has them in a wide range of situations. An experimental study of the validity of this choice is reported together with the design of a letter recognition system. Using a sample of 15 complete 62 symbol alphabets, 13 characteristics were selected. The resulting system recognized correctly 81.9% of the letters presented to it.
Philip Lewis (Thu,) studied this question.