At a conference in Beijing last Thursday, Google representatives announced a brand new technology developed by the company called VisualRank. The new innovation is in essence an algorithm for combining image recognition technology with traditional methods for ranking online images.

According to the New York Times, VisualRank will go far beyond the current method of ranking images in Google’s image search service, and will endeavor to group together images that look alike, and rank them appropriately in the search results

 The traditional method of ranking images by Google and other search engines is to use contextual cues in the form of keywords and image tags to determine what the image represents. But the new VisualRank system will actually scan the picture with image recognition technology, and combine this information with surrounding text data and image tags to determine the topic and relative ranking of the image in search results.

The technology of image recognition has been around for decades now, but has proven very unreliable in detecting faces, for example, and other complicated subjects. But the technology is slowly improving, and Google believes they have developed a method whereby image recognition can be weighed as an important factor in ranking digital images.

Google claims that a trial conducted to test the new VisualRank system resulted in search results that filtered out 83% more irrelevant images, a significant improvement by any standard. Primarily, the image recognition software will identify shapes and common objects, and filter out images that fall too far outside of search term parameters.

But although Google representatives are optimistic about the capabilities of the new system, other industry experts are skeptical. Munjal Shah, the CEO of Riya, believes that what Google is attempting is questionable at best, saying “I think what they’re trying to accomplish is largely impossible.”

Mr. Shah may have a point; the size of the World Wide Web is so incredibly vast that attempting to use image recognition software on every picture online would be a Herculean task — if it were possible at all. And there are other issues as well. For example, when scanning a photograph of a bicycle, image recognition software might be successful if it is a near-field, side-view photo, but would be less likely to accurately categorize the photo if the bicycle was in motion, or the photograph was taken from above.

These kinds of issues raise questions about the usefulness of Google’s VisualRank system, and some digital image experts believe that current image recognition software is likely to mis-categorize 50% or more of the images available online.