An Ensemble of Deep Support Vector Machines for Image Categorization

This paper presents the deep support vector machine (D-SVM) inspired by the increasing popularity of deep belief networks for image recognition. Our deep SVM trains an SVM in the standard way and then uses the kernel activations of support vectors as inputs for training another SVM at the next layer. In this way, instead of the normal linear combination of kernel activations, we can create non-linear combinations of kernel activations on prototype examples. Furthermore, we combine different descriptors in an ensemble of deep SVMs where the product rule is used for combining probability estimates of the different classifiers. We have performed experiments on 20 classes from the Caltech object database and 10 classes from the Corel dataset. The results show that our ensemble of deep SVMs significantly outperforms the naive approach that combines all descriptors directly in a very large single input vector for an SVM. Furthermore, our ensemble of D-SVMs achieves an accuracy of 95.2% on the Corel dataset with 10 classes, which is the best performance reported in literature until now.