Mountain View, California-based Internet giant Google is looking to improve how searchers find images using  its leading search engine, using new methods two of the company's scientists presented in a paper Thursday at a technology conference in Beijing. Google researchers Yushi Jing and Shumeet Baluja outlined a formula called VisualRank which they claim may eventually help searchers using Google find the most relevant images by recognizing the objects contained within, helping to augment an image search system presently based largely on manually typed descriptions. Automating A Cumbersome Tagging Legacy In the paper "PageRank for Product Image Search," which Baluja and Jing presented at last week's International World Wide Web conference, the two offer new methods based on previous research Google has conducted on improving image search and on some of the techniques company co-founders Sergey Brin and Larry Page developed a decade ago. "We conducted a series of experiments based on the task of retrieving images for 2000 of the most popular products queries," Jing and Baluja wrote describing the paper. "Our experimental results show significant improvement, in terms of user satisfaction and relevancy, in comparison to the most recent Google Image Search results," the two noted, referring to Google's existing image search mechanisms, which are based on text entered specifically to describe an image or on text adjacent to an image on a Web page. While Google has made inroads with software recognizing text in images and video as witnessed by its January patent filing with World Intellectual Property Organization, and has along with other firms improved the ability to find faces inside images, determining all the other things contained in images has proven much more difficult, a situation Google hopes to overcome using the techniques Jing and Baluja describe in their paper. Google Research Looks To Improve Image Searching Baluja hopes Google can use the image analysis methods he and Jing researched, and which a team of some 150 Google employees are working on, to improve Web image searches.  "We wanted to incorporate all of the stuff that is happening in computer vision and put it in a Web framework," Baluja noted. " The researchers and their team developed a system for determining the most relevant images for various searches, which resulted in 83 percent less unrelated images shown than in the present Google Image Search system, according to the report. "The ability to reduce the number of irrelevant images shown is extremely important not only for the task of image ranking for image retrieval applications, but also for applications in which only a tiny set of images must be selected from a very large set of candidates," Jing and Baluja wrote in the report. Because of the complex nature of image analysis and comparison the Google team experimented with a small representational subset of the company's vast database of images on the Web, which it refers to as the "most comprehensive image search on the Web," and chose to focus on 2000 of the most searched-for commercial products within the company's Google Product Search service. Combining Image Descriptor Content With Actual Image Contents The new relevancy ranking methods described by Baluja and Jing are meant to include any existing text information describing images, and not supplant it, however the bulk of the methodology described in their paper revolves around image analysis, including such factors as related images and how to prevent people from trying to game the Google image ranking system with images which are not entirely what they claim to be. "Images related to people and celebrities may rely on face recognition/similarity, images related products may use local descriptors, other images such as landscapes, may more heavily rely on color information, etc.," Jing and Baluja noted. The two placed importance with "determining the performance of the system under adversarial conditions," and noted that "it may be possible to bias the search results simply by putting many duplicate images into our index." "We need to explore the performance of our algorithm under such conditions," Jing and Baluja added. By combining these methods with other image detection software such as the January patent for recognizing text within images and video, Google may be likely to someday read much more into images than is presently being done. "Each image is associated with keyword search terms, for example, derived from an image caption, image metadata, text within a predefined proximity of the image, or manual input. Additionally, image search application can include the text extracted from within the images to identify keywords associated with the image. Thus, the text within the image itself can be used as a search parameter," the Google WIPO patent application filed in January noted. Google did not disclose when the methods presented in Jing and Baluja's paper might be implemented. Related Links:
| 
11:32 pm on April 28, 2008 (utc 0)
