digitizationblog News on digitization in libraries and allied institutions

I tried to post this as a comment but found out I couldn't include images....

I just did a quick test of imgSeek based on some cartoons from a collection hosted at the library where I work. Based on this simple test, I'd say imgSeek is pretty accurate:

Query 1 (by example -- I have an image and I want to find others similar to it):

Results:

Query 2 (by drawing -- using a simple pen tool, I drew some crude shapes):

Results:

To respond to Jan's question about applications for imgSeek (or other similar implementations), I'd say that content-based searching is analogous to full-text searching in documents. As you can see from this test, the first image in the results list was pretty close but some of the others weren't. However, I find that it is easier to judge quickly whether a returned image is "relevant" simply because you see the entire thing at once. In contrast, determining if a document is relevant to a given query often requires you to skim a considerable amount of its text.

It's great to see a readily available, open source implementation of content-based image searching. The technology has been around for a while (ex., IBM’s QBIC -- Query By Image Content) but imgSeek packages it up pretty nicely and is very easy to use. To see IBM's technolgy in action, try the Hermitage Museum's online image search.