| 摘要 |
PURPOSE:To extract a feature effective for identification and to realize decision with high accuracy, by reducing noise in a partial area by applying local matching by a partial pattern, and considering a difference between the categories of the local pattern or a character pattern by discriminant analysis. CONSTITUTION:The titled system is the one in which the feature is extracted by performing similarity calculation from linear unification by using the discriminant analysis in every layer, and the localization, for example, shown in 1000, of a stored feature vector group is performed from the pattern memory 10 of a pattern storage means, and it is sent to a linear unifying arithmetic part 1, and a linear unifying matrix in a first layer is read out from a linear unifying matrix memory 30 via a multiplexer 54, then, linear unification is performed. A deflection correcting vector in the first layer is read out from a deflection correcting vector memory 40 via a multiplexer 55, then, deflective correction is performed. After completing the deflective correction, a similarity arithmetic part 3 reads out local feature vectors in order from local feature vector memories 1100, 1101,... via a multiplexer 53, then, similarity can be taken. Obtained similarity is written on a feature vector arranging memory 20.
|
