In this paper we study how Support Vector Machines (SVMs) can be applied to image classification. To enhance classification accuracy, we normalize SVM margins and apply variance reduction techniques to SVM pairwise classification results. From empirical study on a fifteen-category diversified image set, we show that combining SVMs and variance reduction is an effective approach for image classification. This study is a critical step for our on-going effort on the development of a comprehensive approach, closely adapted to SVMs, to image classification.
