Another possibility is to scramble the image names for each user based on a randomly generated key, and create links from the generated image names to the real images. Then you delete the links when the form is submitted, and also run a clean-up routine every so often to get rid of extra links.

The problem with this method is it isn't as elegant as generating a single image on the fly, and perhaps not as efficient either if the images are very small and easily merged.

Or you could dispense with the human segment entirely and base your pass-fail on how long it takes for the form to be submitted. An automatic script will probably submit in under a second, while even the fastest person is going to take at least a couple seconds to type his response.