In image retrieval with relevance feedback, the strategy employed by the system for selecting the images presented to the user at every feedback round has a strong effect on the transfer of information between the user and the system. Using SVMs, we put forward a new active learning selection strategy that minimizes redundancy between the images presented to the user and takes into account assumptions that are specific to the retrieval setting. Experiments on several image databases confirm the attractiveness of this selection strategy. We also find that insensitivity to the scale of the data is a desirable property for the SVMs employed as learners in relevance feedback and we show how to obtain such insensitivity by the use of specific kernel functions.