Integrating multi-features by multiple kernel learning to better classify images

Most recent methods for image classification focus on how to formulate different types of features effectively in a uniform formula. Although these features take on different importance for image classification, most previous work gives the same weight to the features when they are combined. In this paper, we propose an approach to integrate multi-features by following the multiple kernel learning (MKL) framework. By using distinct kernels, we propose to combine different similarity measures for each feature type, that is, the feature fusion is calculated at kernel-level. We employ the SimpleMKL algorithm to solve the MKL problem. As illustrated in the experiments on the images extracted from Corel, Caltech-101 and Flickr 18, our approach outperforms the usual fusion schemes in terms of prediction accuracy.