VisualRank: PageRank adapted to image search
http://research.google.com/pubs/pub34634.html
The famed PageRank algorithm has changed the way search engines return the most relevant web pages for a query. A new area of study, however, is how the idea of distributing importance, or PageRank, among nodes in a graph can be applied to image search. This adapted algorithm is called VisualRank and is described in the linked research paper. As the paper describes, previous image searches have relied purely on image metadata—the content of the page on which the image appears, the hyperlinks that exist to those images, etc—in order to rank the results of an image search. This method has changed with the development of VisualRank.
The paper describes how Google has integrated new research in order to find similarities between images and has incorporated this into the results of an image search. There are a variety of complex ways to measure image similarity, of which the article describes a few—Harris corners, Scale Invariant Feature Transform, Shape Context, and Spin Images. Once these similarities have been calculated, edges and connections can be drawn between homogenous images, or images that share similar content. For instance, two images that both feature the Mona Lisa would be similar and likely to have a connection. As images share more connections among each other, their importance, and thus VisualRank, increases.
From the computed VisualRanks and graph structure, clusters of similar images can be identified and this can be used in various ways. For one, the search engine can choose to feature a few images from each cluster. This allows the user to see a wide variety of results so that they are more likely to see what they were searching for. The VisualRank can also be used to tell which images are most likely the intended topic, similar to how PageRank distributes rank.
In summary, this paper presents an interesting discussion of how PageRank can be applied to image search.