Interactive Image-Retrieval by Estimating Distance Measure

Abstract

A method is described for interactively estimating a distance measure in a feature space for image retrieval. The similarity of images is assumed to be represented by Mahalanobis distance. This method estimates the parameters for the Mahalanobis distance between images in a database and the image the user wants. A system implemented the method presents several images generated or selected by the method and the user selects those that are similar to the one that is to be retrieved. The system again presents images based on the selection. After this cycle of presentation and selection, the distance parameters in the feature space are estimated. The system then retrieves images from the database based on the estimated Mahalanobis distance. The proposed method was experimentally shown to be able to retrieve images without using particular keywards or key images.