Handwritten digit recognition based on a neural SVM combination

In this paper, we present a new multi-class method for handwritten digit recognition that is based on cascade combination of neural networks and support vector machines (SVMs). Binary SVMs are used to provide high linear separation between classes while the neural network generates automatic multi-class decision. In addition, we introduce Jaccard distance into negative distance (ND) SVM kernel. The performance evaluation of the proposed method is conducted on the well-known US Postal Service database. Experimental results indicate that it can significantly reduce the runtime while giving at least the same performance as conventional multi-class SVM methods. Besides, the use of Jaccard distance improves significantly the performance of the standard ND kernel.