Abstract or Summary:
When a user logs on to a recognition based authentication system, he or she is presented with a number of images, one of which is their pass image and the others are distractors. The user must recognise and select their own image to enter the system. If any of the distractors is too similar to the target, the user is likely to become confused and may well choose a distractor by mistake.

It is simple for humans to rule on image similarity but such a labour intensive approach hinders the wider uptake of these mechanisms. Automating image similarity detection is a challenging problem but somewhat easier when the images being used are minimal image types such as hand drawn doodles and Mikons constructed using a computer tool.

We have developed an algorithm, which has been reported earlier, to automatically detect if two doodle images are similar. This paper reports a new experiment to discover the amount of similarity in collections of doodles and Mikons, from a human perspective. This information is used to improve the algorithm and confirm that it also works well with Mikons.