Find Similar

Posted by Grant Mongardi on Tue, May 12, 2009 @ 05:21 PM
Presently, Xinet and many other DAM systems offer reasonably
elaborate ways of finding images based upon both user-supplied
metadata, as well as information provided by the applications
used to create the images. Much of this searching is reliant
upon some level of understanding and expertise by the end user.
This method of searching is generally only as good as the people
both adding the metadata, as well as the folks doing the search.
In other words, the search method itself is very dependent upon
people for it's accuracy.

That may change sometime in the future. Although I have no
expectations that searching user-applied metadata will ever go
away entirely, I do suspect that other options will be forth-
coming that will allow end-users to search based upon the
characteristics of an image. Most images have some sort of
general theme, be it either subject matter, layout, color or
texture. If you can determine these characteristics algorith-
mically, or some combinations of heuristic and algorithmic
techniques, then it would be possible to search initially on
metadata (or simple browse to a particular image to start with)
and then select a defined mechanism to find "similar images"
within some threshold to the reference image. Additionally,
there is also the capability that you can simply upload an
image from your desktop, either existing artwork, or just some
reasonable facsimile drawing that you produce, and let the
search engine find stuff that is similar!

There are already a few of these search engines out there. They
are in very early stages of development, so you can't really use
these as a definition of what's to come, but the results that
they produce are interesting to say the least. From the time
that I've spent looking at these, my suspicions are that most
are initially using a combination of information applied to the
images, either via information provided by the page that they
are placed within, or some sort of metadata/tags provided to
describe the images. From that point, the display of images
offers some mechanism to find images that have characteristics
similar to the reference image - some sort of "find similar"
link or button.

The benefit to this sort of search capability for the typical
DAM administrator is that although metadata searches with still
need to be available, designers will then be able to "find
similar" imagery based upon a texture, or color theme, or
content layout, or even some combination of those with perhaps
some specific metadata, such as "royalty-free". Given the fact
that most designers think and work visually, this will make
reuse of imagery much more productive, and will allow studios
to make the most of their asset library.

If you'd like to see some of these engines in action, or if you
are interested in learning more about the science behind them,
I've provided some links at the end of this article. However,
I wouldn't expect to see these anytime soon on your DAM server,
regardless of who makes it. The technology behind this sort of
searching is truly in it's infancy, and my expectation would be
that it will still need a lot of fine-tuning before we'll ever
see it in a commercial environment such as Xinet provides.
Although, one should note that the fact that Xinet's system
creates and stores previews of a large array of filetypes, this
then adds the capability for one to search across those preview
images rather than just the original image. That adds a layer
of capability that will allow the end-user to not just compare
actual original images, but to compare the previews of all of
the supported filetypes. That means you can find an InDesign,
Quark, or even Word document with similar characteristics to
what could be a reference PSD, or even Vector-based AI file.
Stay tuned, as the future of image similarity searching does
look bright!

Similarity search engines:

Academic Papers on Similarity Searching:

Tags: Xinet, Xinet WebNative Portal, DAM Systems, NAPC blog, Searching with DAM