Fast Multiresolution Image Querying

[danambroseUK]

Fast Multiresolution Image Querying
I have found a paper on the above topic, in short it is a method of using images to search for other images in a database. You draw a quick sketch - ala Paint; it then searchs the database for simlar images.

Obviously there is some clever maths in there which allow this to work... A brief explaination...
http://grail.cs.washington.edu/projects/query/

The full paper
http://grail.cs.washington.edu/projects/query/mrquery.pdf
Has anyone used this method before? Has it been implimented in perl before? There are few implimentatins, the imgSeek program is a great example of this very powerfull stuff.

Anyone like to code this up in perl??

Dan

Edited by Chady -- linkified links.

[merlyn]

What part of "this" do you want in Perl? I'd imagine there's some heavy math for the pixel wrangling, and I wouldn't want to do that in Perl. And the GUI interface will likely require some kit like Tk. Or are you talking about querying their database programmatically through the web?

[danambroseUK]

I intend to make this a web based application. So I would just want perl to do the hardwork encoding the photos and doing the math part...

But the math is alittle full on - way above my level!

Dan

[Chady]

<plug shame="0" perl-related="no">

I have implemented an application that uses this CBIR method. It also uses the Enlightenment Foundation Libraries.

It's still in the alpha stages, and the code is pretty ugly, since it was my first time coding an application in C, but it's open source, so if anyone's interested to have a look and send me critic and comments.

Web: http://emages.chady.net/

</plug>

[anonymized user 468275]

The method for matching a scanned image seems adequately explained in the paper and could be translated from theory into perl.

However, call me skeptical but I just don't believe the implication in the paper that the same techniques can reliably produce the stated results for a painted query image. Any credible attempt to interpret such an image would need to transform the data in terms of human perception and the suggested paper has not even gone in the right direction for this. Such modelling belongs more properly in the field of psychology and psychotherapy, but can describe the concepts the program needs to aim at when transforming the painted query image and reference image for comparison. See 'Metaphors in Mind: Transformation through symbolic modelling' - http://www.alibris.com/search/detail.cfm?S=R&isbn=0953875105&qsort=p

[danambroseUK]

Have a look at http://imgseek.python-hosting.com

Download this and give it a try - I did and the results were impressive on my photo collection.

Your correct; By drawing a cat, your not searching for a cat, simply images which are similar to your drawing. One could build upon this with the addition of meta data & reference images. Which would add some context to the searches and could actually be very powerful.

There are other ways to look at, being shape/path recognition etc. Using the discussed technique could also have a "Find Similar Images" feature- which for say an online photo gallery would be really neat.