Summarizing image collections

Abstract

The widespread use of digital cameras and mobile phones, along with a rapidly growth of image sharing, challenges current image retrieval techniques. It is difficult for image retrieval system to find the semantic meaning of images based on human’s subjectivity and the size of current image databases makes it difficult to organize and search images. This thesis shows that information retrieval techniques can be used to reduce the search space of existing image collections, by creating collection summaries that holds only the most representative metadata from existing image collections. The representative metadata are metadata that are most distinguishing for the specific image collection. The system take advantage of the metadata available for images, which includes user provided tags, date/time, GPS coordinates, and metadata augmented by the system using already available metadata. Higher level representations of user provided terms are located and ensures that the system captures the most descriptive properties of the image collections. The system designed is able to produce collection summaries that capture properties of an image collection that support human’s natural perception as long as enough metadata are available. Also, the system increase the contextual understanding of images as long as date/time (of capture) and GPS coordinates is available for all or some of the images in the collection. The evaluation of the system indicates that grouping user provided terms into higher level representations is very useful for capturing the most important properties of an image collection. Also the evaluation expresses the usefulness of augmenting images with additional metadata and converting numeric metadata into readable terms. The comparisons made in the evaluation indicates that not only are the collection summaries similar to individual test users perception of image collections, but the summaries also includes more descriptive information of relevance.