For the preparation of a tree structure recognition dictionary, in order to obtain the best decision tree, the maximum depth or the sum of the depths should be taken into account. For this purpose, there is provided a system and method for preparing a recognition dictionary which makes it possible to shorten both the maximum recognition time and the total recognition time by successively allocating to nodes of the decision tree features, for which a feature evaluation measure called "maximum estimated depth" for reducing the maximum depth or "sum of estimated node numbers" for reducing the sum of the depths is the smallest.