Personal authentication system and personal authentication method

A personal authentication method comprising:

• classifying a plurality of feature amounts of face image data, which are recorded in advance, into classes;
• calculating as a mapping matrix, coefficients of a linear discriminant equation which uses said recorded plurality of feature amounts as variables;
• extracting a first feature amount from first face image data;
• extracting a second feature amount from second face image data;
• transforming said first feature amount into a first transformed feature amount by using said mapping matrix;
• transforming said second feature amount into a second transformed feature amount by using said mapping matrix;
• calculating as a similarity, a normalized correlation value between said first transformed feature amount and said second transformed feature amount;
• and
• judging that said first face image data and said second face image data are image data of same person when said similarity exceeds a predetermined threshold,
• wherein said calculating said mapping matrix includes:
• calculating a first normalized correlation value between a first average vector of feature amounts belonging to a first class and a first feature amount belonging to said first class;
• calculating a second normalized correlation value between a second average vector of feature amounts in a second class and said first feature amount, wherein said second normalized correlation value is a maximum value of normalized correlation values between said first feature amount and average vectors of feature amounts in respective classes other than said first class;
• and
• updating said mapping matrix by using a steepest descent method such that a variation of a loss with respect to said mapping matrix approaches zero, wherein said loss is based on said first normalized correlation value and said second normalized correlation value, and
• said transforming includes transforming a feature amount extracted in said extracting into a transformed feature amount by using said updated mapping matrix.

A computer-readable storage medium as a non-transitory tangible medium, storing a personal authentication program that causes a computer to execute a personal authentication method comprising:

• classifying a plurality of feature amounts of face image data, which are recorded in advance, into classes;
• calculating as a mapping matrix, coefficients of a linear discriminant equation which uses said recorded plurality of feature amounts as variables;
• extracting a first feature amount from first face image data;
• extracting a second feature amount from second face image data;
• transforming said first feature amount into a first transformed feature amount by using said mapping matrix;
• transforming said second feature amount into a second transformed feature amount by using said mapping matrix;
• calculating as a similarity, a normalized correlation value between said first transformed feature amount and said second transformed feature amount;
• and
• judging that said first face image data and said second face image data are image data of same person when said similarity exceeds a predetermined threshold,
• wherein said calculating said mapping matrix includes:
• calculating a first normalized correlation value between a first average vector of feature amounts belonging to a first class and a first feature amount belonging to said first class;
• calculating a second normalized correlation value between a second average vector of feature amounts in a second class and said first feature amount, wherein said second normalized correlation value is a maximum value of normalized correlation values between said first feature amount and average vectors of feature amounts in respective classes other than said first class;
• and
• updating said mapping matrix by using a steepest descent method such that a variation of a loss with respect to said mapping matrix approaches zero, wherein said loss is based on said first normalized correlation value and said second normalized correlation value, and
• said transforming includes transforming a feature amount extracted in said extracting into a transformed feature amount by using said updated mapping matrix.