RI: III: Small: IInterlinking Image Collections

Stanford University, Stanford CA

Investigators

Abstract

The availability of digital cameras, improved networking, and the diminishing cost of memory has made it easy to capture, share, and store large image collections. With dense sampling, there are many connections and correlations among these captured images, as they effectively record the same or visually similar objects. This project aims to build such networks of linked images on a large scale, store the inter-image relationships in the form of a graph or simplicial complexes called Image Webs, study the properties of these networks, and exploit them for a variety of applications. Establishing links between parts of images based on image content analysis, and doing so on the scale of millions of images, is a computationally demanding task. Since it is impractical to do this for all image pairs, techniques are developed for attempting to establish links only between pairs for which (1) a link is likely to exist and, (2) the link adds substantially to what is already known about the connectivity of a particular Web. A deeper understanding of the global structure of image webs as topological complexes can aid this link prediction process. Methods are developed for effectively navigating these large structures and for constructing useful maps over them. Integration with more symbolic information associated with images is possible by transferring information around in this vast network. The project is of a highly interdisciplinary nature, combining techniques from continuous applied mathematics, traditionally used in signal processing and image analysis, with methods from discrete mathematics and network theory.