In the practical applications of content based image retrieval (CBIR), ordinary users normally have no expertise in constructing a high quality query, so some noisy examples will present. The noisy examples may be irrelevant images mislabeled by user or weak relevant images which can not well represent the set of relevant images. Both of them will degrade the retrieval performance significantly. However, most existing CBIR schemes failed to address this problem. In this paper, we introduce noise tolerant mechanism into biased discriminant analysis (BDA), which is an effective CBIR scheme especially in the case of small number of examples. First, an actual image is selected as the center of positive examples rather than the average point which is sensitive to noisy examples. Then, original BDA is extended by allowing to associate with each example a probability that it is labeled correctly. The extensive experiments carried out on a Corel image collection show that the proposed scheme has good noise tolerant ability that can improve the retrieval performance of original BDA significantly under the noisy circumstance.