This document discusses a new approach for content-based image retrieval using relevance feedback and support vector machines. The proposed Biased Maximum Margin Analysis (BMMA) and semi-supervised BMMA (SemiBMMA) techniques aim to overcome two drawbacks of existing SVM-based relevance feedback methods. First, they treat positive and negative feedback differently since they have distinct properties. Second, most existing techniques do not utilize information from unlabeled samples. The BMMA differentiates positive from negative feedback locally, while the SemiBMMA incorporates unlabeled data using a Laplacian regularizer. Experiments on a large image database show the proposed techniques improve performance over standard SVM relevance feedback approaches.