A voting principle of multiple features for Chinese character recognition system using neural network classifiers

Abstract

We propose a modified SCONN (self creating and organising neural network) classifier (MSC), which uses the algorithm of learning vector quantization. We adopt two commonly used features, namely the crossing-count feature and contour-direction feature in our recognition system. The experimental results show that MSC performs well and has advantages of being simple in network structure and efficient in computation time. A voting principle useful in selecting candidates based on measurement values derived from variable error distance is proposed. We test several formulas for calculating the confidence level (ballots) of candidates, and show that the proposed voting principle can increase up to 10% in recognition accuracy than otherwise using the MSC alone.