Google engineers last week presented an interesting paper at the WWW2008 conference in Beijing which proposes to apply its PageRank system of finding relevant Web pages to radically improve the accuracy of image search results using Google. This new technology is being called VisualRank, according to an fascinating story on the subject in the New York Times. The paper, titled “PageRank for Product Image Search,” (PDF) was published by Yushi Jing and Shumeet Baluja of Google. In it they talk about using PageRank to analyze the “visual link structure” that can be created among a group of similar images. This paper proposes to move away from the current model of many image search engine rankings, described as using “the text clues of the pages in which images are embedded to rank images.”
The new model would “identify ‘authority’ nodes on an inferred visual similarity graph and propose an algorithm to analyze the visual link structure that can be created among a group of images.” A numerical weight would be assigned to each image and, according to the paper, ranking would occur based upon “expected user behavior given the visual similarities of the images to be ranked.” The assumption, this blog points out, is that “people are more likely to go from an image to other similar images.” The Google engineers ran this model through some experiments using 2000 of the most popular products queries. They found “significant improvement, in terms of user satisfaction and relevancy, in comparison to the most recent Google Image Search results.”
Source: PC World, Martin
