PURPOSE: To obtain a feature set having the reduced redundancy by calculating a recognition rate based on the feature obtained from a relative feature evaluation function.
CONSTITUTION: An initial feature selecting part 1 selects a feature that is most effective to identification out of a multi-dimensional feature vector and uses the feature as the first one of a selected feature set. A 1st feature evaluating part 2 selects the candidate features out of the feature vectors excluding the element of the selected feature set. A 2nd feature evaluating part 5 generates the candidates of a new selected feature set based on those candidate features selected by the part 2. A recognition rate calculating part 4 measures a recognition rate based on the generated candidate feature. A selected feature set deciding part 5 decides a new selected feature set out of those candidates. Then a processing end deciding part 6 decides whether the recognizing result obtained by using the new selected feature set satisfies the desired value or not. If not, the part 2 works again to perform the similar processing.
KIDA HIROMI