An user preference information based kernel for SVM active learning in content-based image retrieval

Relevance feedback is a critical component for content-based retrieval systems. Effective learning algorithms are needed to accurately and quickly capture the user's query concept, under the daunting challenges of high dimensional data and small number of training samples. It has been shown that support vector machines (SVMs) can be used to conduct effective relevance feedback in content-based image retrieval. Most recent work along these lines has focused on how to customize SVM classification for the particular problem of interest. However, not much attention has been to paid to the design of novel kernel functions specifically tailored for relevance feedback problems and traditional kernels have been directly used in these applications. In this paper, we propose an approach to derive an information divergence based kernel given the user's preference. Our proposed kernel function naturally takes into account the statistics of the data that is available during relevance feedback for the purpose of discriminating between relevant and non-relevant images. Experiments show that the new kernel achieves significantly higher (about $17%) retrieval accuracy than the standard radial basis function (RBF) kernel, and can thus become a valid alternative to traditional kernels for SVM-based active learning in relevance feedback applications