Pattern recognition apparatus for determining a reference pattern by K-L expansion utilizing a power method

Claim 1

A method of obtaining a reference pattern from a plurality of input patterns of a specific category, comprising the steps of: a first step of obtaining a covariance matrix K of said plurality of input patterns of the specific category, said covariance matrix K having a elements in n columns and n rows; a second step of multiplying said covariance matrix K by a first vector ξ1 consisting of elements of a first column of said covariance matrix to obtain a second vector ξ2; a third step of obtaining a first maximum element of aboslute value ABS(B1) in said second vector ξ2; a fourth step of multiplying said covariance matrix by said second vector ξ2, obtained in said second step, to obtain a third vector ξ3; a fifth step of obtaining a second maximum element or absolute value ABS(B2) of said third vector ξ3; a sixth step of determining convergence of said second vector and said third vector obtained in said second and fourth steps, respectively, said convergence being determined by obtaining a first value and a second value from the first maximum element of absolute value ABS(B1) and from the second maximum element of absolute value ABS(B2) which are obtained in said third and fifth steps, respectively, and comparing the first value with the second value; a seventh step of repeatedly performing calculations in said second and fourth steps within a predetermined calculation time, if the convergence of said second vector and said third vector does not occur in said sixth step; an eighth step of storing said vector obtained in said fourth step as an eigen vector of a first axis and said maximum element in absolute value of said vector as an eigdn value of said first axis in a directionary memory, when the convergence is determined in said sixth step; a ninth step of obtaining a new covariance matrix K' adapted for obtaining eigen vectors and eigen values of axes subsequent to the first axis, by subtracting the eigen vector and eigen value of the first axis from said covariance matrix K; a tenth step of determining whether eigen vectors and eigen values for a predetermined number of axes have been obtained; and an eleventh step, a power method step, of sequentially obtaining eigen vectors and eigen values up to the predetermined number of axes by repeating steps corresponding to said second to tenth steps with respect to said new covariance matrix K', when it is determined in said tenth step that the eigen vectors and eigen values have not been obtained up to the predetermined number of axes.

Claim 2

A method of obtaining a reference pattern from a plurality of input patterns of a specific category, comprising the steps of: a first step of obtaining a covariance matrix K of said plurality of input patterns of the specific category, said covariance matrix K having a elements in n columns and n rows; a second step of multiplying said covariance matrix K by a first vector ξ1 consisting of elements of a first column of said covariance matrix to obtain a second vector ξ2; a third step of obtaining a first maximum element of aboslute value ABS(B1) in said second vector ξ2; a fourth step of multiplying said covariance matrix by said second vector ξ2, obtained in said second step, to obtain a third vector ξ3; a fifth step of obtaining a second maximum element or absolute value ABS(B2) of said third vector ξ3; a sixth step of determining convergence of said second vector and said third vector obtained in said second and fourth steps, respectively, said convergence being determined by obtaining a first value and a second value from the first maximum element of absolute value ABS(B1) and from the second maximum element of absolute value ABS(B2) which are obtained in said third and fifth steps, respectively, and comparing the first value with the second value; a seventh step of repeatedly performing calculations in said second and fourth steps within a predetermined calculation time, if the convergence of said second vector and said third vector does not occur in said sixth step; an eighth step of storing said vector obtained in said fourth step as an eigen vector of a first axis and said maximum element in absolute value of said vector as an eigdn value of said first axis in a directionary memory, when the convergence is determined in said sixth step; a ninth step of obtaining a new covariance matrix K' adapted for obtaining eigen vectors and eigen values of axes subsequent to the first axis, by subtracting the eigen vector and eigen value of the first axis from said covariance matrix K; a tenth step of determining whether eigen vectors and eigen values for a predetermined number of axes have been obtained; and an eleventh step, a power method step, of sequentially obtaining eigen vectors and eigen values up to the predetermined number of axes by repeating steps corresponding to said second to tenth steps with respect to said new covariance matrix K', when it is determined in said tenth step that the eigen vectors and eigen values have not been obtained up to the predetermined number of axes.