PURPOSE: To evenly discriminate a pattern with a high precision by using the statistical value of each discrimination value to determine each inputted precedent pattern and generating data of a neural network to exclude the subjectivity of a discriminator.
CONSTITUTION: Reference patterns to which respective precedent patterns inputted through a precedent pattern input device A belong are discriminated in a discrimination value input device C by plural discriminators, and the discrimination value is inputted for each discriminator. Consequently, plural discrimination values are inputted for each precedent pattern, and the degree of variance of discrimination values is generates learning data E. Thus, synthetic data of plural discriminations is learnt by the neural network, and an influence of the individual difference among discriminators and erroneous discrimination is reduced. The neural network is used to obtain the result of discrimination G of a newly inputted pattern to be discriminated as if plural dlscriminators discriminated this pattern to obtain the degree of variance of plural discrimination values.
MATSUMOTO HIDEAKI
SHINONAGA HIROYUKI
OKASHITA HIROSHI
KOYAMA TAKESHI
SUZUKI EIICHI
KOBE STEEL LTD