Applying relevance feedback to a photo archival system

Abstract

Relevance feedback techniques have been successfully applied to document retrieval systems (DRS) to refine queries and document representations. This paper represents a first attempt to apply and optimise relevance feedback techniques to improve retrieval in a text-based image archival system. The design exploits the rapid assessment possible with image data to facilitate query refinement and collect large amounts of relevance feedback data. This data can then be used to extend incomplete image descriptions, thus ameliorating problems associated with text annotation of images. Visually oriented relevance feedback and query modification is implemented using direct manipulation of icons. An algorithm designed for dynamic modification of image descriptions based on relevance feedback is proposed, implemented, and experimentally tested. Initial experiments show significant improvements and demonstrate the potential of using these techniques for image retrieval applications.