Method for optimizing a recognition dictionary, so that similar pattern can be distinguished better

摘要

A discriminant function is defined by conventional learning discriminant analysis (22) and a value of the discriminant function is calculated (23) for all the training patterns in the in-category pattern set of each category and for all the training patterns in the rival pattern set of the category. The in-category pattern set is composed of all the training patterns defined as belonging to the category. The rival pattern set is composed of the training patterns that belong to other categories and that are incorrectly recognized as belonging to the category. An in-category pattern subset and a rival pattern subset are then formed (24) for each category. The in-category pattern subset for the category is formed by selecting a predetermined number of the training patterns that belong to the in-category pattern set and that, among the training patterns that belong to the in-category pattern set, have the largest values of the discriminant function. The rival pattern subset for the category is formed by selecting a predetermined number of the training patterns that belong to the rival pattern set of the category and that, among the training patterns that belong to the rival pattern set, have the smallest values of the discriminant function. A linear discriminant analysis operation is then performed (25) on the in-category pattern subset and the rival pattern subset to obtain parameters defining a new discriminant function. The reference vector and weighting vector stored in the recognition dictionary for the category are then modified using the parameters defining the new discriminant function.