Inferring image semantics from collection information

Abstract

There is a continuing desire and need for improving the processes of describing and searching for digital images. While good progress has been made adapting traditional information retrieval techniques to perform these tasks, processing images still presents a number of challenges not encountered when working with just text. This project implements a system allowing for the indexing and searching of collections of images where images are not individually described. Such collections are typically largely unsuitable data material for traditional text-based image search systems. The basic idea underlying this system is to not search for images directly, but instead first search for the collections they are part of. When a set of relevant collections have been found, one can then apply the collection-level semantics to the images belonging to these collections. An important concept in this project is that one image can be a part of several collections. Such images can be identified when collection information is added to the system, allowing one to link the same image to several collections before searches are run. Semantic information from several different collections can then be applied to the same image, potentially giving a better idea of what the image actually depicts.