PURPOSE: To perform pattern recognition with high identification rate even when sample data is too small or too large by providing a lower limit value WMIN and an upper limit value WMAX at weight data W used in an identification function.
CONSTITUTION: For example, a handwritten character is read by an image scanner, and a feature vector S is generated at a feature extraction circuit 10. A generated feature vector S is sent to a recognition decision part 20, and the feature vectors Ri-Rj of a reference character registered on a dictionary part 30 and their weight vectors Wi-Wj are compared with the feature vector S. The feature vector Rj of a reference pattern nearest to the feature vector S of an input pattern is discriminated by using the identification function provided with the weight data Wij=f(σij)(where, σij: standard deviation around γij, f(σij): arbitrary decreasing function expressed in WMIN≤f(σij)≤WMAX(0≤WMIN <WMAX<∞) in (σij)≥0).
MACHIYA SHUKO