Image Recognition and Web Search
Life in the 21st century gives us wonderful new technology that allows us to shorten our attention spans and lower our persistence and determination for fact finding. If I don’t know something, I google it. If I don’t find what I’m looking for within a half a page of Google’s intelligently determined link list, I google something else. Luckily, Google’s page-rank algorithm is talented enough to retrieve whatever information I seek within a few tries (ie. the 3 first links my eyes happen to land on). Page-rank is able to predict fairly quickly and accurately what webpages I am looking for based on a network of pointers and attributes of that web page, but what happens when I am looking for information that doesn’t contain such objects and attributes? Specifically, I am talking about images.
When I google web search the term “Cornell”, I get a list of web pages starting with Cornell’s very own website home page, Cornell’s open-edited Wikipedia page, and the Cornell Athletics Big Red page. Scrolling down, I see Cornell University’s facebook fan page and Cornell news results (Congratulations to Professor Peck on the new CTO Position at NASA). For the entire first page, I see entirely Cornell-related content, and only such relevant content. When I search “Cornell University” in images, I see the glowing clock tower lighting up the evening, beautiful Ithaca autumn foliage and a scene of women dressed in feathers and bikinis on a dance floor. This within the first two rows of images. There is a jarring lag between the effectiveness of image search behind web page search because of the obvious logistic problems with intelligent image recognition in computers.
Google researchers have been working on an algorithm for improving web-based image search. This algorithm is called VisualRank and combines several recognition techniques with network concepts to rank images. Just like the nodal representation of web pages, researchers hope to apply images into a similar network of similar images. Although humans can immediately recognize images that do not fit in search results, it is tedious to do this through Google’s vast catalogue of images. Google attempted to game-ify it’s tagging with a web game that had users match tags on a presented images. This backfired due to confusing game rules and the unfortunate bad nature of many internet users, leaving ranking up to software to algorithmically tag images. Image recognition software combined similar shapes and color schemes to create networks of images. In combination with software, search engines watch the images users view under the assumption that consecutive images are related. Together these properties form a network of images that Google implements an analogous algorithm to PageRank upon as if a network of webpages were inputted. In Google’s paper “PageRank for Product Image Search”, researchers evaluated their combinatory ranking system using images of 2000 of the most popular product queries in Google’s web search and the new system produced 83% less irrelevant images. This is less imperfect and shows that further improvement on the intelligent image recognition front is necessary before I know what to expect out of a google image search.
Source: http://www.nytimes.com/2008/04/28/technology/28google.html and Google Image Results for “Cornell University”