Media Wonk

Link This | Email this | Comments (0)

Google learns to "see" - April 28, 2008

A pair of researchers at Google has published a paper describing a new method for visual search that relies on an algorithm to estimate "visual hyperlinks" between images to produce a final page rank. Although the paper is technical (Media Wonk sucks at math) the basic idea seems to be to use traditional text-based search to collect a gallery of likely images in response to a query, and then performing a second analysis that looks for visual "links" that connect the images to generate the ranking. No actual human-created links exist between the images, but the algorithm can infer them from other clues without having to perform a computationally expensive visual processing of each image. The researchers claim the method produced 83% fewer irrelevant images than traditional tag-based methods.

Google isn't the first company to tackle the problem of visual search. But given its vast search expertise and resources its entry into the visual search sweepstakes is notable. The paper doesn't indicate whether the virtual visual-link method could be applied to video search.

The New York Times has the story here; the full paper is here (PDF).

[Content Protection & Management]  [R&D]   LEAVE A COMMENT