Relevance feedback methods for content-based image retrieval have shown promise in a variety of image database applications. These techniques assume two (relevant and irrelevant) class relevance feedback. While simple computationally, two class relevance feedback often becomes inadequate in providing suÆcient information to help rapidly improve retrieval performance. We propose a locally adaptive technique for content-based image retrieval that enables relevance feedback to take on multi-class form. We estimate a exible multi-class metric for computing retrievals based on Chi-squared distance analysis. As a result, local data distributions can be suÆciently exploited, whereby rapid performance improvement can be achieved. The eÆcacy of our method is validated and compared against other competing techniques using a number of real world data sets.