N-SVM combination and tangent vectors for handwritten alphanumeric character recognition

We propose a system for handwritten alphanumeric character recognition that is based on the learning of tangent similarities. The specific tangent vectors which constitute the a priori knowledge of each class are generated from the training data. Based on these tangent vectors, similarities with respect prototypes of classes are computed to be used as data features. In addition, we investigate the use of N-SVM combination in order to improve the error rate and reduce the runtime compared to the standard SVM. Experiments conducted on a database obtained by combining USPS and C-Cube data indicate that the proposed system gives the best performance in terms of training time and error rate.