Proceedings of the Annual Meeting of the Cognitive Science Society

Completely automated public Turing test to tell humans apart (CAPTCHA) aims to exploit the ability gaps between machines and humans to distinguish between them. However, the rapid development of artificial intelligence technology in the past decade has significantly narrowed the gap in some tasks based on natural images (e.g., object detection and recognition). Mooney images (MIs) are important research materials in the field of cognitive science. Compared to natural images, we perceive MIs relying more on the iteration between feedforward and feedback processes. In this paper, we explored an intriguing question: Can MIs be used to distinguish between machines and humans? Before this study, we first proposed a framework HiMI that generated the high-quality MIs from natural images and also allowed flexible adjustment of the perceived difficulty. Next, we designed two MI-based Turing test tasks related to foreground-background segregation and object recognition, respectively. We compared the performance of human subjects and the deep neural networks on these two tasks. The experimental results indicate the significant gaps between the deep neural networks and humans, providing evidence for the potential of MIs in the design of CAPTCHA schemes. We hope that HiMI will contribute to more research related to MIs in the fields of cognitive science and computer science.