The characteristic selection problem in recognition systems

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