“If a picture is worth a thousand words, how important is the capability to find the just the right image of an object from the entire Web?”
Search engine leader Google Inc., which already has a very successful “PageRank” technology for sifting web pages, the company has now taken a step further by unveiling a cutting-edge image ranking technology for ranking similar images at the International World Wide Web Conference, according to media reports Tuesday.
A couple of Google researchers have developed an algorithm that quantifies the characteristics of an image and rank their relevance to a search term. The technology, which Google is calling “VisualRank,” has applications for e-commerce and marks a great advancement in machine learning, the researchers said.
At the International World Wide Web Conference in Beijing, two Google scientists Shumeet Baluja and Yushi Jing announced the evolution of their search tool that uses advances in computer which manages to perform image recognition and comparison of similar images to determine a ranking which will push to the first search results those pictures which are considered most relevant to a specific search.
The ultimate aim is to train computers to move beyond text into the effective identification of “rich content” — the shapes, colors and context of images that humans recognize with little effort.
Presently, images are graded by analyzing the text near the image and the image’s file name. “We wanted to incorporate all of the stuff that is happening in computer vision and put it in a Web framework,” said Baluja, a senior staff researcher at Google. He was joined in presenting the new technology by fellow Google expert Jing.
VisualRank is placed as PageRank for images. Google’s PageRank helps determine a site’s value, based on content and scaled from 0-10. The higher the PageRank, the higher the site becomes visible in organic search listings for related keywords.
Despite decades of endeavor, image analysis remains a largely unsolved problem in computer science, researchers said. For example, while progress has been made in automatic face detection in images, finding other objects such as mountains or tea pots, which are instantly recognizable to humans, has lagged.
Search online for a specific type of image today and the results will reveal more or less exclusively the text that is associated with them. This latest innovation presented by two Google researchers, however, assures to use visual cues in the images themselves to rank their relevance.
The method in Google’s paper changes that. A group of images recovered for a query using traditional search methods is then further analyzed. Image recognition software finds which images in the group seem most similar to each other. It then estimates “visual hyperlinks” between them to produce a final ranking.
Analyzing a tremendous amount of images in depth requires, quite logically, incredible computing power. To develop VisualRank, they carried out a series of experiments retrieving images for 2,000 of the most popular products queries, including “iPod,” “Xbox” and “Zune”. They later determined the top 10 images from its ranking system, gleaned in part from Google Image Search results.
“Google is endlessly innovating to provide users with the most relevant results possible, and Image Search is no exception,” the company said. “We are very excited about the VisualRank technology. It marks an important step forward in integrating many novel machine-learning, statistical and computer-vision approaches towards providing better image search results to our users.”
The paper presented by the two Google experts is called “PageRank for Product Image Search.”