Abstract
The way images have been labeled has evolved and matured over the past few years. Before, images uploaded to the Internet were manually labeled by the user uploading the initial image. Computer vision techniques have also been implemented to increase the number of labels for any given image; however, the process is still not yet mature. Therefore, the ESP game was designed to make use of humans’ desire to be entertained in exchange for useful work. As players participated in the game, they would simultaneously be labeling images on the Internet. The data collected from the games would then be used to improve image search. However, a shortcoming of the ESP game is that it provides no way of verifying whether or not a label is indeed correct or valid for a given image. Experienced players may collaborate to abuse the game in order to gain higher rankings; as a consequence, invalid or irrelevant labels are generated for images. Thus the data must somehow be filtered. CAPTCHAs (Completely Automated Public Turning test to tell Computers and Humans Apart) were designed to present users with a problem that is easy for humans to answer, but hard for computers to answer. If a user can successfully pass a CAPTCHA test, then we assume the user is indeed human and not an automated script or bot. In this thesis, we develop a new type of CAPTCHA by presenting users with a challenge that helps solve the image verification problem. As a result, we are able to not only verify that the user is human, but also simultaneously verify and rank labels associated with a given image.