PURPOSE: To reduce the number of times for repeating learning and to improve the ability of identification by automatically judging the pattern of combination, which is hardly discriminated, from an evaluation threshold value.

CONSTITUTION: A two-dimensional character pattern such as a printing character, etc., to be recognized is optically read and binarized at a character read processing part 1, and character patterns are extracted one by one at an extraction processing part 2. As a one-dimensional vector sent from the character pattern extraction processing part 2, the character pattern is divided into groups for each similar pattern at a group division processing part 3 and sent to networks 1, 2,...N of respective groups. Then, a character decision processing part 4 decides which character is the recognized result, and a character code corresponding to this character is outputted from a character code output processing part 5.