A Simple and Effective CAPTCHA by Exploiting the Orientation of Sub-images Cropped from Whole-size Photos

Abstract

Automated detection of image orientation has previously been studied as an important problem in intelligent image processing and computer vision. For this problem, numerous methods and tools have been developed by adopting approaches such as objects segmentation, color feature analysis and machine learning e.g., Support Vector Machines(SVMS). But conversely, the difficulty of image orientation can be used to examine the robustness of a CAPTCHA(Completely Automated Public Turing test to Tell Computers and Human Apart). The automated image orientation problem previously only had been studied and solved using typical photos which almost include important semantic cues such as people, bright sky, dark ground and vertical edges. In this paper we propose a simple prototype CAPTCHA, which exploits the hardness of orienting sub-images cropped from a whole digital photo. Our CAPTCHA takes 8 sub-images from base-photos and rotates them randomly. Then we present them to the user, who is required to find the correct orientations of the 8 sub-images. The true orientation is easily obtained since most current high-end digital cameras have an automatic mechanism to store its orientation in EXIF. Thus we can simply and easily obtain the image orientation without applying complicated computation. For our experiment, we have collected about 1850 base photos that provide more than 100,000 different sub-images. Experiment showed that the accuracy of our CAPTCHA with humans is about 95%. We think this sub-image orientation is hard to solve by an automated procedure since all previous machine learning procedures have only considered whole photos with enough semantic cues, rather than partial image segments. Another advantage of our system is that user interaction is simpler(there are four choices) and more intuitive than a common text-based system or the previous image orientation method with arbitrary rotation. Experiment showed that common users performed at most two rotations for each sub-image. The total time to complete orienting the 8 sub-image orientation was less than 15 seconds which is significantly shorter than that of previous image-based CAPTCHAs.