Multi-Class SVMs Based on Fuzzy Integral Mixture for Handwritten Digit Recognition

The major drawback of support vector machines (SVMs) is that the training time grows fastly with respect to the number of training samples. This issue becomes more critical for multi-class problems where a set of binary SVMs must be performed. This is the case of the one-against-all (OAA) approach, which is the most widely used implementation of multi-class SVMs. In this paper, we propose a new divide-and-conquer method to reduce the training time of OAA-based SVMs. Experimental analysis is conducted on handwritten digit recognition task. The results obtained indicate that the proposed scheme allows a significant training and testing time improvement. In addition, a significant improvement in generalization performance was obtained