Relevance feedback methods for content-based image retrieval have shown promise in a variety of image database applications. These techniques assume two-class relevance feedback, relevant and irrelevant. While simple computationally, two-class relevance feedback often becomes inadequate in providing sufficient 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. For each given query, we estimate local feature relevance based on Chi-squared analysis using information provided by multiclass relevance feedback. Local feature relevance is then used to compute a flexible metric that is highly adaptive to query locations. As a result, local data distributions can be sufficiently exploited, whereby rapid performance improvement can be achieved. Experimental results using real image data validate the efficacy of our method.
展开▼