Breaking Visual CAPTCHA

I just ran across this wonderful article by Mori and Malik on a technique to break the EZ-Gimpy visual CAPTCHA images that attempt to distinguish between human and automated blog commenters. (Comment, memberlist, trackback and forum spam are a huge problem for sites like ours, and automated Turing tests like visual CAPTCHA are a promising way to avoid them). Mori and Malik propose a cool lettershape-matching algorithm based on log-polar histogram plots of edge-detected objects found within the image. They further refine their letter detection by tying the strings of possible letter choices together in an attempt to form dictionary words. Their algorithm does better than I was able to do on some of the overlaid word images.