Emerging image generation with flexible control of perceived difficulty

Abstract

Emerging images (EI) are two-tone and contain a number of discrete speckles. If certain speckles are appropriately organized together, we will perceive a meaningful object, which reflects the closed-loop information processing of human visual cognition. EIs hold significant application value. They can be used in studies of perceptual organization in cognitive psychology. Additionally, they can also serve as a CAPTCHA mechanism to distinguish humans from bots in the field of network security. Both applications require a method for generating EIs that can flexibly adjust the perceived difficulty. Although universal style transfer (UST) models are capable of generating images in a specific style, it can be challenging to adjust the generated results to meet different user needs. In this paper, we present a novel EI generation framework and achieve flexible control over the perceived difficulty of the EIs by extracting and quantifying different cognitive cues and setting the corresponding parameters to adjust the proportion of these cues rendered in the EIs. The experimental results both qualitatively and quantitatively demonstrate that our methods generates EIs with higher quality while allowing for more flexible control over the perceived difficulty. Furthermore, we prove the potential of EIs as a CAPTCHA through sufficient experiments.