Off-line recognition of handwritten middle age Persian characters using moment

Abstract

In this paper, the performance of several moment invariant features combined with various classification methods for the recognition of middle age Persian manuscripts is presented. Specifically, Legendre moments (order 2 to 12), Zernike and pseudo-Zernike moments (order 2 to 15), and the set of invariant moments (ϕ1, ϕ2, ..., ϕ7) are used as features. These features are computed from four versions of character images; (1) grayscale character images (Set A), (2) semithresholded character images (Set B), (3) binarized character images (Set C), (4) character skeleton (Set D). For classification, we have used the minimum Mean Distance (MMD), k-nearest neighbor (KNN), and Parzen methods. The experiment yielded a 2.86% error rate (97.14% classification rate) with pseudo-Zernike moments on the semithresholded character images (set B).