Character recognition using statistical moments

Abstract

This paper presents a character recognition system that is implemented using a variety of statistical moments as features. These moments include H u moment invariants, Af fi ne moment invariants and Tsirikoli as-Mertzios moments. Euclidean distance measure, cross correlation and discriminati on cost were used as the classi fication techniques. The mean of the intraclass standard deviations of the features was used as a weighting factor during the classification process to improve recognition accuracy. The system was rigorously tested under different conditions, including using different number of training sets and documents with different fonts. It was found that Tsirikoli as-Mertzios moments with weighted cross correlation classifier provided the best recognition rates. © 1999 Elsevier Science B.V. All rights reserved.