Relevance feedback has attracted the attention of many authors in image retrieval. However, in most work, only positive example has been considered. We think that negative example can be highly useful to better model the user's needs and specificities. In this paper, we introduce a new relevance feedback model that combines positive and negative examples for query processing and refinement. We start by explaining how negative example can help mitigating many problems in image retrieval such as similarity measures definition and feature selection. Then, we propose a new relevance feedback approach that uses positive example to perform generalization and negative example to perform specialization. When the query contains both positive and negative examples, it is processed in two steps. In the first step, only positive example is considered in order to reduce the heterogeneity of the set of images that participate in retrieval. Then, the second step considers the difference between positive and negative examples and acts on the images retained in the first step. Mathematically, the problem is formulated as simultaneously minimizing intra variance of positive and negative examples, and maximizing inter varicance. The proposed algorithm was implemented in our image retrieval system "Atlas" and tested on a collection of 10.000 images. We carried out some performance evaluation and the results were promising.