Injection of human knowledge into the rejection criterion of a neural network classifier

The purpose of unconstrained handwritten numeral recognition is to assign a numeral to one of ten classes or reject it. The challenge is to maintain a high performance and not to misrecognize confusing patterns. In some applications, it is desirable to reject a pattern instead of running the risk of misclassifying it. In order to improve the reliability of a single neural network classifier on confusing numerals, knowledge from five human experts is gathered and analyzed. A new way to construct database and represent the required output values in the output layer of MLP´s training process is given in this paper. Experiments on a synthesized confusing database and a real database show that the proposed approach will facilitate the design of a highly reliable single neural network classifier