Accuracy improvement of handwritten numeral recognition by mirror image learning

This paper proposes a new corrective learning algorithm and evaluates the performance by a handwritten numeral recognition test. The algorithm generates a mirror image of a pattern that belongs to one class of a pair of confusing classes and utilizes it as a learning pattern of the other class. This paper also studies how to extract confusing patterns within a certain margin of a decision boundary to generate enough mirror images, and how to perform an effective mirror image compensation to increase the margin. 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 99.41% respectively in the recognition test for handwritten numeral database IPTP CD-ROM1.