Interesting project that retrieves images solely on the basis of associated text.
"The underlying hypothesis is that, given a set of textual Web nodes linked with a non-textual node, i.e. an image, the text associated with these nodes could be used to represent the content of the non-textual node, since the existence of the links signifies a semantic relationship between these nodes and the non-textual node."
In the near future, there is a graphics format being developed by the W3C called Scalable Vector Graphics [w3.org] (SVG). I understand that, with the SVG standard, direct search for letterforms within vector graphic files may be possible.
That's still no help with raster files, and they won't go away, so this project NFFC reports on has some real potential.
However, vector images could well grow by leaps on the web once a public standard is available, rather than just the proprietary one with MacroMedia Flash.