Clustering with projection distance and pseudo Bayes discriminant function for handwritten numeral recognition

This paper investigates the usage of the projection distance and the pseudo Bayes discriminant function as the distortion measure for handwritten numeral clustering problem. These distortion measures not only refer to the mean vectors but are also related to the covariance matrixes of subclasses, thus, the distribution of subclasses are reflected on the obtained clusters, and the accuracy of recognition can be improved. A series of evaluation experiments are performed on the handwritten numeral database NIST SD3 and SD7. The experimental results show that the recognition rate has been increased from 97.35% to 98.35%, which is one of the highest rates ever reported for the database