JP2945454B

PURPOSE:To curtail processing quantity by selecting the condition of a search part by comparing whether a dispersion parameter of the number of categories divided into two by its branch item is large or small. CONSTITUTION:A feature vector is generated from an input pattern, and the pattern discrimination is executed by referring to a two-division tree table related to a pattern predicted as a candidate category prepared in advance from the feature vector. The two-division tree table is constituted by setting the division as a search condition in the case by a bit of each dimension of a feature vector in which a classification effect of a category assembly is large. In such a state, by comparing whether a dispersion parameter is large or small, in the search part being near the root of the two-division tree table, a bit of a dimensional value of the feature vector which can classify largely, many categories is selected as the search condition, and in the search part being near a leaf of the end, a bit which can discriminate each similar category is selected. In such a way, as for a search path in which an input pattern is inputted and discriminated, only the bit required for discriminating other category is necessary, and all bits are not required, therefore, the calculation quantity is reduced.