[Search-l] vizual searching
Charlie Hull
charlie at juggler.net
Thu Apr 26 16:14:02 UTC 2007
EVANS, KENNETH (WOODY) wrote:
> I'm a librarian, and I've been thinking about visual searching a lot
> because I get frustrated with the searches I have to settle for... If I
> want a "blue" picture, I have to count on someone using the word "blue"
> as a tag, subject heading in metadata, or in the filename
> (blueboat.png).
>
> I've been working on an idea I was calling "Vizibiz" -- a search engine
> that would let you upload or paste images, and return results that were
> visually similar. The notion was that every pixel would be analyzed,
> and metadata would be created for each pixel -- color, position in pic,
> color of neighboring pixels. Vizibiz would also search in traditional
> metadata and filenames, file sizes, and other attributes (page context).
Unfortunately it's considerably harder than this to work out if images
are 'similar' - I've seen a few projects that have had a go, but I don't
think there is anything that actually works. Think of the problems of
magnification, of different RGB colours that actually look similar to
the human eye, of partial (clipped) images...there's a reason that a
large part of our brain is taken up with processing the data from our
retinas :-)
>
> This idea never got off the ground because 1) I don't write code. I'm
> code ignorant. 2) When I started searching for open software that could
> get me started doing this job, I discovered a couple of companies that
> already seem to be on this track:
> http://www.evisionglobal.com/tech/basic_demo.html
> http://able.mulabs.com/
>
> What if this were web-based? And free for anyone to use?
Then everything would be lovely :-) however I suspect that anyone who
has cracked this problem will be making money selling it to security
agencies and governments, rather than giving it away to the world. Don't
hold your breath.
Charlie
>Is there any traction for this sort of search interface in Search
Wikia? Kind of
> like a "music genome project", but automated and for visual rather than
> audio elements...
>
> Thoughts?
>
More information about the Search-l
mailing list