A Classifier of Similar Characters using Compound Mahalanobis Function based on Difference Subspace

Abstract

To distinguish similar characters, it is preferable to construct a classifier using a projective feature space which differentiates two similar categories. The classifier CMF has been proposed for a discriminant function, in similar characters recognition. In the CMF, a subspace is constructed by some eigenvectors, that corresponds to the smallest eigenvalues, is applied as projective feature space. A difference vector of two class-mean feature vectors are assumed as the difference between two similar categories, the CMF is constructed by projecting a feature vector onto this difference vector. In this paper, we propose new discriminant function expanding the CMF. In proposed method, we treat the Difference Subspace, which is difference between two subspaces as difference between two similar categories. The efficiency of the proposed new discriminant function has been demonstrated in similar characters recognition through extensive experiments on hand-written Japanese characters derived from the ETL9B database.