Mirror image learning for handwritten numeral recognition

Abstract

This paper proposes a new corrective learning algorithm and evaluates the performance by handwritten numeral recognition test. The algorithm generates a mirror image of a pattern which belongs to one class of a pair of confusing classes and utilizes it as a learning pattern of the other class. Statistical pattern recognition techniques generally assume that the density function and the parameters of each class are only dependent on the sample of the class. The mirror image learning algorithm enlarges the learning sample of each class by mirror image patterns of other classes and enables us to achieve higher recognition accuracy with small learning sample. Recognition accuracies of the minimum distance classifier and the projection distance method were improved from 93.17% to 98.38% and from 99.11% to 99.37% respectively in the recognition test for handwritten numeral database IPTP CD-ROM1 [1]. © Springer-Verlag Berlin Heidelberg 2001.