Google researchers, Shumeet Baluja and Yushi Jing, presented a paper titled ‘PageRank for product image search’ at the international World Wide Web Conference held in Beijing last week. They say that their new technology will help improve image search results, and these results will be more relevant to customer needs.
At present search engines rank images based on the presence of keywords in text associated with the image, rather than the image itself. Results are often unsatisfactory and inappropriate. Image analysis results remain poor in spite of decades of research in the field.
The New York Times reports that search geniuses intend to introduce a system for image searches that is similar to PageRank, which is used for Web search. The new PageRank for product image search, or ‘Visual Rank’ as it is called, will actually try to understand the picture or analyze the visual link structure.
Image recognition, which is expensive and time consuming, usually breaks down with any images other than faces, alphanumeric characters and very specific objects. The new technology will instead try to identify more objects and also perform a link analysis with image processing. In this system, a numerical weight will be allotted to all images, thus weighing its relative importance vis-Ã -vis other images.
Classical image recognition methods can only compare images with a known image. However, with VisualRank, Google will use “visual themes”, to rank each image from a given set of images, based on how well it matches the theme. For example, when the word McDonald’s is used, the algorithm would identify the famous golden arches from their logo as the theme. A picture where the arches are prominently displayed would then be ranked higher than one in which the arches are in the background.
During a recent study Baluja, Jing and 150 panelists surveyed 2,000 of the most popular products search on Google. It was found that the VisualRank algorithm reduced irrelevant results by 83%. As of now all this has been done only on an experimental level and it remains to be seen whther Google can scale the technology and apply it to their entire database of images.