Kernel based image classification

Abstract

In this study, we consider low-level image classification, with several machine learning algorithms adapted to high dimension problems: kernel-based algorithms. The first is Support Vector Machines (SVM), the second is Bayes Point Machines (BPM).We compare these algorithms based on strong mathematical results and nice geometrical arguments in a feature space to the simplest algorithm we could imagine working on the same representation. We use different low-level data, experimenting lowlevel preprocessing, including spatial information. Our results suggest that the kernel representation is more important than the algorithms used (at least for this task). It is a positive result because it exists much more simpler and faster algorithms than SVM. Our additive low-level preprocessings only improved success rate by few percents.