An Extended Learning Vector Quantization Algorithm Aiming at Recognition-Based Character Segmentation

Recognition-based segmentation strategies have greatly improved the performance of optical character recognition systems. The key issue of these strategies is to design a classifier that can provide accurate rejection information. Many learning algorithms, such as GLVQ and H2M-LVQ, are not suitable for large category sets and multiple prototypes. More seriously, they often suffer from local minimum state and overtraining. In this paper, we propose an extended learning vector quantization algorithm which can efficiently train the nearest prototype classifier with negative samples. The cost function is based on multiple confusable prototype-pairs so that our algorithm is insensitive to initialization. We also introduce the criterion of safe zone to avoid overtraining. Experimental results show that the classifier trained by our proposed method can achieve good recognition performance and can provide accurate rejection information for segmentation.