PURPOSE: To read characters with less erroneous read, by preserving original features, which are ignored in stages of the character pre-processing conversion and the recognized feature extraction, and using these original features in the final discrimination for a series of characters.

CONSTITUTION: When a series of characters 8 are inputted, a cutting part 1 cuts characters 8 into every one character to output a character pattern 9 for every one character. An original feature extracting part 2 takes out, for example, height information of every one character and stores it in a storage part 6. In a pre-processing conversion part 3, the size in the longitudinal direction and the lateral direction is normalized in every one character to obtain a normalized pattern 10. This pattern 10 is supplied to a recognizing part 4, and recognized features are extracted in the recognizing part 4 and are collated with contents of a dictionary 5 to obtain an answer 11 for every one character. An editing part 7 searches characters having the same answer in answers 11 for individual character stored in the storage part 6. Original features having the same answers are compared with each other; and if their difference is larger than a preliminarily set reference value, these characters are rejected.