Improved handwritten character recognition using second-order information from training set

The problem of improving the capability of statistical character classifiers based on finite and sparse training sets is addressed. A significant improvement is obtained coupling standard classifiers based on the k-nearest neighbours technique with a second higher level classification stage. This method has been applied to three existing classifiers reducing the error rate at zero rejection of ~ 17%. © 1993, The Institution of Electrical Engineers. All rights reserved.