Has it not happened that you use Google search to search for images of say Danio rerio (a well-known fish). You get 16 900 hits. Unfortunately it is all too impossible to find a fish on many of those photos, because Google finds pages with the text Danio rerio having an image. And it takes time to browse through 16 900 images. This kind of image search is useful for finding the unexpected, such as anatomical drawings, phylogenetic trees, and people working on zebrafish, and definitely worth trying now and then. Without the image filter, you get 1 510 000 hits for Danio rerio, and that is just too much to browse in the hope of finding an image of the fish.
Google Labs now have a new tool under development that looks promising. It is called Similar Images, and uses some kind of image pattern recognition. What you do hear is that you specify a search for images as usual, e.g., Danio rerio, get the same result (but 16 500), but some images have the link “similar images”. Click one of those, and you get a subset of the 16 500, and on succeeding searches using the same method, you get down to 300-500 look-alike images, and most likely you have exactly what you want within reach.
This might be a tool for matching photos of unidentified fish specimens, and could also be helpful to check fish identifications on the web. Maybe could be used for other things than fish? Seems to work well with Paris Hilton (a well-known human), but for what …?