PURPOSE: To extract features necessary for the rough classification and discrimination of an input original character pattern stably and efficiently, by providing a pattern buffer memory, feature extracting circuits, registers, rough classification dictionary, and collation arithmetic circuit.

CONSTITUTION: The original input character pattern inputted through a photoelectric conversion part and quantized into a binary signal is stored in a pattern buffer memory 1 temporarily. Pieces of information read out of the pattern buffer memory 1 by making a longitudinal scan are inputted to an LU and LD extracting circuit 2, PU and PD extracting circuit 3, and CU and CD extracting circuit 4 respectively to extract pieces of information. Pieces of information read out of the memory 1 by making a lateral scan are inputted to an LL and LR extracting circuit 5, PL and PR extracting circuit 6, and CL and CR extracting circuit 7 to extract pieces of information. The pieces of information obtained by the circuits 2W7 are set in registers 8W10 and classified roughly. The collation arithmetic circuit 11 compares said pieces of information with pieces of dictionary information registered in the rough classification dictionary 12 to decide on a category which satisfies coincidence condition within a specific range.