Click here for more
Computing news and columns
* * * * *
Visual tool focuses search
Ditto.com looks for Internet imagesBy Andrew Zajac
Nov. 15, 1999
CHICAGO - In concept, the Internet is a miracle, a vast trove of information ready to be mined with a brief dance of the fingers across a keypad. But the reality, as any user knows, is considerably grimmer.
Just try finding exactly what you need in that dense, uncharted jungle of data.
Naperville, Ill.-based software developer Michael Lyons stared at the online thicket and saw opportunity.
"I looked at the Internet and said, `We're about to enter the greatest renaissance since the 14th century,' and I wanted to participate. Search is a fundamental element. If you want people to participate, you have to be able to find stuff."
So Lyons founded ditto.com to develop and market a visual search engine, focusing on the relatively underserved niche of locating photographs and other images on the World Wide Web.
The 2 1/2 -year-old company provides visual search for Snap, the C/NET-NBC portal, and is looking for deals with other high-profile Web sites as it refines its approach in what is a notoriously quirky branch of Internet technology.
The sheer enormousness of the World Wide Web - an estimated 800 million pages and growing by the minute - defies any plausible hope of systematic indexing. The most comprehensive text search engine, Northern Light, roots through a meager 16 percent of the Web.
But that's just the beginning of search challenges. Search engines aren't yet smart enough to deal with the common sense problem. Words like "port" or "bridge" have multiple meanings and can be either a noun or a verb. It all depends on a mammoth amount of context, far more than can be mustered by any search engine designed so far.
Natural language search technology tries to resolve this ambivalence by analyzing how a word is used in a sentence. But this kind of search engine - you can test one on the Ask Jeeves or Google sites - is far from perfect.
Web designers keen for a prominent ranking of their sites sometimes increase the incidence and prominence of the key words that search engines look for, in effect exaggerating the content of pages and tainting the quality of search. Other site builders try to fool Web crawlers with their metatags - hidden key words meant to convey an idea of a page's content - or leave out metatags.
It's no easier on the other side of the computer screen, where consumers fumble with whether and when to employ connectors like "and" and "or" on a particular search engine and whether or not to use quotation marks on another - all painful evidence of a lack of a simple, standardized protocol.
Visual search starts with text search, so it's plagued with all of the latter's shortcomings, plus others inherent in the fledgling state of image-recognition technology, which analyzes patterns of pixels and color to try to "read" a picture.
"Right now search engines do a poor job, and visual search is slightly worse than text search," said Paul Hagen, who covers search technology for Forrester Research of Cambridge, Mass.
But to ditto.com's Lyons, 48, being among the first to market beats waiting for a highly refined product.
"I don't have time to wait for the perfect technology," Lyons said, noting that "the big media companies are coming into this space" and it gets hugely more expensive to build a market identity with each passing year.
Indeed, the two other competing visual search engines, Scour.net and Alta Vista's Picture Search, are owned by investors headed by Hollywood mogul Michael Ovitz and Internet investment company CMGI Inc. of Andover, Mass., respectively.
Ditto.com will pair pictures requested by customers with photos of related merchandise. "What we are about to do is create the impulse buy on the Web," he said.
A search for Arnold Palmer, for example, would produce photos of the golfer and, in a separate section of screen, photo-based ads for clubs, gloves or balls.
Lyons said the Internet is ripe for an alternative to the banner ad, which he dismissed as less effective than direct snail mail and cluttered with too much message that makes it too hard to buy.
Virtual merchandising is a head-turning proposition, said Bob Goldstein, a Los Angeles-based analyst with Future Image, who wrote a recent report on visual search.
"The direct linking of search results with being able to buy products is one of the directions of the future," Goldstein said.
There's a critical mass of would-be shoppers assembling on the Internet, but "people are having trouble finding things. One of the keys to e-commerce is being able to find stuff," Goldstein said.
Ditto.com has $10.4 million in backing from Swiss-based venture group European Technology and Finance and is looking for $15 million more specifically for branding and advertising. Ideally, that new money would come from a U.S.-based venture firm and from media companies that control valuable proprietary picture content.
Ditto.com employs 30 people - 23 in Naperville and seven at a research facility in Emeryville, Calif.
Its founder was an Army brat who collected a political science degree from Utah State University, an MBA from Boston University and the rank of captain from the Army, where he learned applications development.
(Ditto.com was built by programmers mainly from Northwestern University and the University of Chicago working under chief technology officer Steven Schmidt.)
Lyons says that what makes ditto.com's searches more effective than other systems' is that the engine is constantly creating and refining a database of image links and does not try to do Web searching on the fly.
Ditto.com begins its work by sending out a search agent, or spider, that crawls the Web to round up the logical suspects: files that have telltale endings like .gif and .jpeg, which denote images.
Lyons said ditto.com looks through about 8 percent of the Web on its periodic forays, focusing its search on subjects in the news and on what search engine users are asking for, based on an analysis of consumer query patterns.
Then Ditto uses image recognition software to sort out its catch. It distinguishes between black and white and color images, between illustrations and photographs, people and things, and it weeds out pornography, banner ads and other detritus.
"People are the hardest ones in terms of getting it right. It can tell the difference between a picture and a drawing, but between three people it's going to have a tough time," Lyons said.
Smut still is the Internet's most successful business. "Forty percent of what we pick up is porn," Lyons said.
All told, the image recognition filter rejects 96 percent of the 60 million or so images initially gathered, leaving a cache of about 2.5 million pictures.
Next, software reads the text surrounding photos and ranks them for relevance. For instance, Sammy Sosa in the heading of a picture caption would give it a higher relevance ranking than the slugger's name buried in the body of the caption.
Pictures then are shrunk to thumbnail size, so that when ditto.com executes a search, several images can quickly be displayed on the screen.
Finally, the database of downsized photos and links moves to 55 part-time human reviewers, who fish out any remaining porn and further rank photos of the same subject by trying to answer the question: "If I were a consumer, which pictures would I want to see?"
"It's our position that the only way you're going to get great relevancy is to combine technology with human filters," Lyons said. "The key is lower numbers and higher relevancy."