Mutual Information Based Kernel for Image Classification

Significant research has been done to derive kernel function for the linear learning machines like Support Vector Machines (SVM). Kernel methods provides a way to add generative model advantages into a discriminative model. Kullback - Leibler (KL) divergence-based kernel is an excellent example of this combination. In this paper, we proposed new Mutual Information kernel that is an alternative kernel of the KL divergence kernel. We compared our proposed kernel with KL-divergence based kernel and some other standard kernels like Poly, RBF, and Sigmoid kernels. We have done comparison of these kernels with the image classification task. We have also described two different image representations to obtain a generative model for the use of KL-divergence or MI-based kernels.