PURPOSE: To exactly recognize a character-string containing a voiced sound mark and a semivoiced sound mark, by providing a subcharacter pattern extracting part for coupling a minute subcharacter pattern to an adjacent character pattern.

CONSTITUTION: An image inputted from an image input part 1 is binary-coded and stored in an image memory 2, and thereafter, a character-string R whose absolute position is determined in advance is segmented. Subsequently, in a subcharacter pattern extracting part 4, a histogram is derived with respect to a character-string segmented by a rectangle R, and a subcharacter pattern is segmented. Also, in the subcharacter pattern, a minute subcharacter pattern whose height and width are smaller than a reference value is extracted, and based on an interval to the subcharacter pattern adjacent to this minute subcharacter pattern, this minute subcharacter pattern is coupled to the adjacent subcharacter pattern. Moreover, the pattern whose character width satisfies the reference value is determined as an individual character pattern.