Google Reports Gains in Web Searches for Digital Images

By John Markoff

Researchers at Google say they have developed a software technology intended to do for digital images on the Web what the company’s original PageRank software did for searches of Web pages.

On Thursday, at the International World Wide Web Conference in Beijing, two scientists from Google presented a paper describing what the researchers called VisualRank, an algorithm for blending image-recognition software methods with techniques for weighting and ranking images that look most similar to what the searcher has in mind.

Although image search has become popular on commercial search engines, results are usually generated by using cues from the text that is associated with each image.

Despite decades of effort, image analysis remains a largely unsolved problem in computer science, the researchers said. For example, while progress has been made in automatic face detection in images, finding other objects, like mountains or tea pots, which are instantly recognizable to humans, has lagged behind.

"We wanted to incorporate all of the stuff that is happening in computer vision and put it in a Web framework," said Shumeet Baluja, a senior staff researcher at Google, who made the presentation with Yushi Jing, another Google researcher. The company’s expertise in creating vast graphs that weigh "nodes" or Web pages based on their "authority" can be applied to images that are the most representative of a particular query, he said.

The research paper, "PageRank for Product Image Search," is focused on a subset of the images that Google, the search engine business, has catalogued because of the tremendous computing costs required to analyze and compare images. To do this for all the images indexed by the search engine would be impractical, the researchers said.

Google does not disclose how many images it has catalogued, but it asserts that its Google Image Search is the "most comprehensive image search on the Web."

The company said that in its research it had concentrated on the 2,000 most popular product queries on Google’s product search, words like iPod, Xbox, and Zune. They then sorted the top 10 images both from their ranking system and the standard Google Image Search results.

With a team of 150 Google employees, they created a scoring system for image "relevance." The researchers said the retrieval led to an 83 percent reduction in irrelevant images.

Google is not the first to make a foray into the visual product search category. Riya, a Silicon Valley start-up, introduced Like.com in 2006. The service, which refers users to shopping sites, makes it possible for a Web surfer to select a particular visual attribute, like a certain style of brown shoes or a style of buckle, and then be presented with similar products available from competing Web merchants.

Rather than relying on a text query, the service focuses on the ability to match shapes or objects that it might be hard to describe in writing, said Munjal Shah, the chief executive of Riya.

"I think what they’re trying to accomplish is largely impossible," he said. "Our belief is there is not large-scale solutions."

Shah said there had been a number of technology demonstrations by Google Labs researchers, including a 2005 project that used machine learning techniques to recognize the sex of a person in an image. But the company has been slow to deploy its research, he said.