1 Introduction

To get acquainted with the contents of large image databases, when the right annotation is missing, comes down to some form of browsing or relevance feed-back approach. For an overview of content-based image retrieval methods and systems see the special issue of [ 1 ]. In this paper we restrict our attention to browsing methods. Because browsing thumb-nails may take days for large image databases, one way of speeding-up the visual inspection would be to show the thumbnail frames in rapid succession. We therefore considered using the thumb-nails as video frames. However if the consecutive frames of the video are taken from thumb-nails at random, the video display speed must be kept low (a few frames per second) for comfortable viewing. In real video, frames appear typically at a rate of 25-30 per second, a speed which is only acceptable because of the large average frame to frame coherence where scene breaks do not occur in rapid succession. If one could sort the thumb-nails into a maximal coherent sequence, the display of this sequence can be speeded-up to a higher number of frames per second without enlarging the chance of missing a potentially interesting image. To help sort the database images into a display order with high frame to frame coherence, we have tried to exploit the full image to image distance matrix measures available from our precalculated Query by Example similarity matching methods. Two methods have been implemented to generate a shortest path using the distance measures.