Google trumpets PageRank for pics Nearly a decade ago, Google unveiled an algorithm called PageRank, reinventing the way we search for web pages. Now, the company says, it has a technology that can do much the same for online image search. Last week, at the International World Wide Web Conference in Beijing, two Google-affiliated researchers presented a paper …
When I do google searches for images I often find I get quite a few identical images of different sizes and quality. From the sounds of it this system would identify a good picture of say, Paris, and then lots of copies of it. I hope there will be a way of filtering potentially duplicate images and only displaying the best resolutioned (sp?) one. At the moment just changing 'large' 'medium' 'small' is not entirely helpful.
You know in reality there picture rank is going to consist of little more that analysing the pallete of the picture and if it has a large percentage of flesh tones rank it higher up the list.
Thus image search will be better at finding porn.
This method also works at big university's where users home directorys are world readable allowing you to find all the mucky stuff quite quickly :)
When I was younger, I recall a TV program put out by PBS about artificial intelligence which included a segment on a US Military program in the 80s to train a computer to visually recognize Soviet Union tanks. The researches did the old training bit, showing the system lots of photos of Russian and 'Merican tanks and telling the system which was which. Then after a suitable training period, they fed the system photographs it had never seen before and asked it to identify the units. It failed miserably, at a higher than random rate. When an audit was performed to determine what the problem might have been, it came to light that all the photos it flagged (incorrectly or correctly) as Russian had trees in there along with the tank. Further investigation lead to the revelation that all the training photos of Russian tanks had trees in them, but the 'Merican tanks did not.
The point is, if you use pattern recognition systems, you're not likely to know what pattern the computer will decide is important.
While some pattern recognition systems are "black box", not all are. As AC above says, most image analysis software is _very_ unsophisticated in the features it uses, but it can still perform surprisingly well.
In classifier systems the most important element by miles is the the features you use as input. The classifier (neural net vs support vector machine vs decision tree etc) can't turn a sow's ear into a silk purse.
Paris, because a home video was leaked onto the internet showing her engaging in various sexual acts, including sexual intercourse and fellatio, with a man called Rick Salomon.
The Register 
Biting the hand that feeds IT
Copyright. All rights reserved © 1998–2024
