Implicit image tagging via image-based CAPTCHAs

Abstract

In order to facilitate image retrieval, we proposed architecture of a collaborative image tagging system, which takes the form of image-based CAPTCHAs. The system includes methods of using current image recognition algorithms to extract ‘target objects (objects to be tagged)’ from images, identifying ‘target attributes’ of the extracted target objects (an image tagging system), differentiating between humans and computers (as CAPTCHAs), and generating verified test images (a test image generator). The proposed image tagging system was applied to recognition of human faces, which can be automatically detected by current image recognition algorithms, and evaluated its applicability in two steps. In the first step, we implemented GenCAPTCHA, which was designed to identify the gender of human faces. We concluded that GenCAPTCHA reliably identified gender-indiscernible faces because the consistency of our user responses was 89.05%. A single, eight-image GenCAPTCHA challenge was completed in 12.41 seconds (average), with a human success rate of 86.51%. This success rate could be increased by filtering error-prone test images. In contrast, the probability of passing a GenCAPTCHA challenge by random guessing was 0.006%. In the second step, we implemented AgeCAPTCHA to tag human faces according to age category. In contrast to gender, some age groups were difficult to clearly distinguish, so we applied a two-layered response structure that uses the indistinguishable categories, not to differentiate between humans and computers, but to tag the images. In our user studies, the user responses showed high consistency (80.19%). The median completion time of a single AgeCAPTCHA challenge was 15.75 seconds, and the human success rate was at least 88.21%. From all these results, we conclude that our image tagging system is robust enough to resist machine attacks, and easy enough for human users, to be used in practice.