Discrimination of similar characters with Quadratic Compound Mahalanobis Function

CMF is an efficient function in distinguishing a character from similar ones. This distinction is performed by correcting the Mahalanobis distance using feature vectors projected onto a certain subspace. However, this approach may have some limitations for the projective space, because the decision boundary surface of two similar characters is generally very complicated, There is also a method to estimate the decision boundary surface of two similar characters using nonlinear conversion, but the degree of variance-covariance matrix used in this estimation is very large. In this paper, we propose a CMF applied quadratic nonlinear conversion. We refer to our proposed method as the Quadratic Compound Mahalanobis Function (QCMF). It is shown that the QCMF can discriminate similar characters with a smaller amount of computation than discrimination with quadratic nonlinear conversion, and that in recognition experiments with similar characters, the accuracy of a recognition system with QCMF is better than that with CMF.