I like the idea in theory, but there's a few of practical issues

1) I got the circle, and I tried 8 times before giving up. I was pretty damn close but it just denied me. There seems to be more lenience in the straight line options, perhaps just stick to them.

2) What's the fallback for blind users?

3) There are 16 images that are used, and you can have as many guesses as you want. I'm guessing that it's easily brute forced. How about dynamically generate a shape from n lines?

This seems to be easily bypassed even if you dynamically generate the shapes.

What stops me from submitting the form with curl?

doesn't seem to work too well in this: http://i.imgur.com/bCLVN.png

Innovative, new idea but i don't think it is a great one. Bots can easily read pixels and pass that rest, don't you think?