| 摘要 |
PROBLEM TO BE SOLVED: To provide an arithmetic device capable of calculating a dimension-reducing matrix A suitable for pattern recognition, even when a distribution of a feature vector shows multiple peaks. SOLUTION: This device acquires a sample of the feature vector of the pattern, belonging to the class for each of the classes, in executing pattern recognition (S110). With respect to each of the classes, a degree of similarity G between the vector (X<SB>ij1</SB>, X<SB>ij2</SB>) is calculated (S150) for each of the combinations (X<SB>ij1</SB>, X<SB>ij2</SB>) of the feature vector (the sample). Specifically, the degree of similarity G is calculated so that the larger the distance D=¾¾X<SB>ij1</SB>-X<SB>ij2</SB>¾¾ is such that the smaller the real value (non-negative) is, the smaller the distance D=¾¾X<SB>ij1</SB>-X<SB>ij2</SB>¾¾ is, and the larger the real value is taken. Through weight assignment by this degree of similarity G, a covariance matrix B<SP>Local</SP>between local classes and a local class covariance matrix C<SB>i</SB>are calculated (S160, S170). Furthermore, the dimension reduction matrix A, in which a ratio of the covariance matrix B<SP>Local</SP>between the local classes with a generalized average of the local class covariance matrix C<SB>i</SB>becomes a maximum, is calculated (S180). COPYRIGHT: (C)2011,JPO&INPIT
|
