With respect to two pattern sets obtained on different conditions, a
feature-extraction matrix, which maximizes between-class scatter and
minimizes within-class scatter, is found respectively. A first feature
amount is calculated using one of the feature-extraction matrices. The
first feature amount and the two matrices are retained in a referential
database. A pattern is determined to a second feature amount--extracted by
applying another feature-extraction matrix to a pattern input--by
extracting a most similar element out of the first feature amount retained
in the referential database.