experiment

On Thursday, Google Research engineers presented a paper at the International World Wide Web Conference in Beijing on PageRank for Google Images (pdf) to improve search results for photos, art and graphics. The system promises better image results than are currently available when searching in Google Images and may eventually improve Google Universal Search SERPs.

Key takeaways: Google’s breakthrough uses the “wisdom of the crowd” and contextual signals to rank the relevancy of images. VisualRank does not improve on a search engine’s ability to identify people or determine activities in a photo. The biggest benefits will be reduction of duplicate image content in search results and reduction of “image spam” or inappropriately tagged photos.

This morning The NY Times reported on the new ranking algorithm that identifies and analyzes “authority†nodes and “visual link structure” between a group of images. As with PageRank, images are assigned numbers to define their relevance and relative importance.

Google conducted a series of experiments by retrieving images for 2,000 of the most popular products queries in Google. Users in the experiments were more satisfied by the results and felt they were more relevant.

The Google SERP image shown here displays top ranking results for a group of queries. You can judge for yourself how intuitive and relevant the results are. Google notes an interesting result for the query “Picasso Paintingsâ€; not only are all the images by Picasso, one of his most famous, “Guernicaâ€, was selected first.

We’re assuming the search queries related to the product, “Febreze” were spelled correctly, unlike the typo in the paper misspelled as: “Fabreze.” The current Google image results for keyword “fabreze” are quite different.

Winners: Trademark owners of big brands and commercial products

Losers: All those people who spent innumerable hours tagging photos in Google Image Labeler:

All-time Top Google Image Labeler Contributors
1. SunChaser has 22,961,020 points
2. Zip has second with 22,353,450 points
3. FrD AUTO no car has 15,460,240 points
4. MC DUDE no man has a close 15,350,830 points
5. Mighty is hot on the heels of MC DUDE: 15,339,300 points

Google believes a Web page author will likely choose relevant images for a topic. People, though, don’t typically link to content based on the relevance of images. People link to text.

Google gives an example of an ambiguous query (McDonalds) with a logo that can be identified in photos that link commercial searches. That’s a “visual theme†or ‘visual signal” among all the photos. There may be lots of other themes that can define the relative “strength” of common and commercial images.

In terms of overall performance on queries, the proposed VisualRank displayed fewer irrelevant images than Google for 762 queries. Google’s standard image search producee better results in only 70 search results. In the remaining 202 queries, both approaches tied. Google notes in the majority of these queries, there were no irrelevant images).