Multi-class relevance feedback content-based image retrieval

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 classes. While simple computationally, two-class relevance feedback often becomes inadequate in providing sufficient information to help rapidly improve retrieval performance. In this paper we propose a multi-class form of relevance feedback retrieval to try to exploit multi-class information. For a given query, we use a χ2 analysis to determine the local relevance of each feature dimension with multi-class relevance feedback. This information is then used to customize the retrieval metric to rank images. By exploiting multi-class information, our method is able to create flexible metrics that better capture user perceived similarity. In a number of image data sets, the method achieves a higher level of precision with fewer iterations, demonstrating the potential for substantial improvements over two-class relevance feedback retrieval.